Re: RFC GDB Linux Awareness analysis

[Peter Griffin <peter.griffin@linaro.org>]

Hi Andreas,

On Thu, 20 Aug 2015, Andreas Arnez wrote:

Many thanks for your feedback.
> 
> Sorry for the late reply.  Yao Qi pointed me to this RFC; I've just
> missed it before.  For those who have missed it as well, the original
> posting was here:
> 
>   https://sourceware.org/ml/gdb-patches/2015-06/msg00040.html
> 
> Independently of that proposal I have presented some thoughts about
> improving GDB's Linux kernel debugging support on the GNU Tools
> Cauldron.  Surprisingly or not, some of the goals look very similar.
> The slides are here:
> 
>
>https://gcc.gnu.org/wiki/cauldron2015?action=AttachFile&do=view&target=Andreas+Arnez_+Debugging+Linux+kernel+dumps+with+GDB.pdf

Thanks for the link to your slides, that is very intresting. From your slides
LKD implements exactly some of the features you were proposing.

For example info threads in LKD reports running and sleeping threads
and the loadable module support, resuses solib infrastructure inside GDB.

> 
> In that talk I was focusing more on kernel dump debugging rather than
> live debugging, but I also tried to emphasize the fact that the feature
> sets required for both (should) have a large overlap.  Only a few
> features are specific to one or the other, e.g. live debugging requires
> a notification method for refreshing the task list and the modules list.
> 
> From the resonance in the talk I got the impression that there is indeed
> wide-spread interest in having improved kernel debug support in GDB.

Ok that is good news, one reason for posting the RFC to the mailing
list was to try and gauge what the GDB community support was for having such a
feature. So it is good to know at least so far it seems positive.


> So I believe posting the patches to get more feedback would be worth-while.

Ok, so the existing LKD code is on quite an old GDB version (7.6). Also the whole
LKD patchset is currently quite large (15k LoC). So my plan was to try and 
reduce this as much as possible by removing parts of LKD that can (or already
have) python implementations e.g. dmesg.

Essentially the aim would be to try and reduce the LKD patchset as much as possible
to the "core features" which I consider to be task awareness and loadable module
support. At this point we could look at what might need adding
to the GDB python API to migrate even more parts into python.

Posting the whole patchset currently would involve porting to the latest GDB
version and before doing that I'd like to get a better idea of what parts are
likely to need re-writing or at least only ports the necessary parts.

> 
> On Wed, Jun 03 2015, Peter Griffin wrote:
> 
> > Hi GDB community,
> >
> > Overview
> > ========
> >
> > The purpose of this email is to describe a useful feature which has been developed
> > by STMicroelectronics inside of GDB for debugging Linux kernels. I will cover at
> > a high level some of the implementation details and what I see as the advantages
> > / disadvantages of the implementation. I will also cover some other alternative 
> > approaches that I'm aware of.
> >
> > The purpose is to facilitate discussion with the GDB experts on this
> > mailing list as to what the "correct" way to implement this functionality would
> > be.
> 
> I'm copying gdb@sourceware.org instead of gdb-patches, since I think
> it's better suited for now -- until we are actually discussing a patch.

Thanks :)

> 
> > The end goal is to have an upstream implementation of this functionality.
> >
> > Introduction
> > ============
> >
> > STMicroelectronics has a patchset on top of vanilla GDB which adds much better
> > Linux kernel awareness into GDB. They have called this GDB extension LKD (Linux
> > Kernel Debugger). This GDB extension is primarily used in conjunction with ST's
> > JTAG debugger for debugging ARM / SH4 SoCs, and is implemented as an internal
> > GDB extension, written in C.
> >
> > The LKD extension is nicely abstracted from the underlying JTAG interface,
> > and I have used it to debug a ARMv7 kernel running in QEMU via gdbremote.
> >
> > ST would like to contribute these patches back to GDB, as we think it could
> > be useful not only for ARM Linux debugging, but also other CPU/OS combinations
> > in the future.
> >
> > LKD can be broadly split into the following parts: -
> >
> > 1) Linux task awareness
> > 2) Loadable module support
> > 3) Various Linux helper commands
> >
> > The next section looks at each of the three parts, and any other implementations
> > I'm aware of that currently exist.
> >
> >
> > 1) Task Awareness
> > =================
> >
> > 1.1) LKD Linux Task awareness
> > =============================
> >
> > When using mainline GDB for debugging a Linux kernel via JTAG GDB will typically
> > only show the actual hardware threads.
> >
> > The LKD task awareness extension (lkd-process.c) adds the ability for GDB to
> > parse some kernel data structures, so it can build a thread list of all kernel
> > threads in the running Linux kernel.
> >
> > When halting the processor (via a breakpoint, ctrl+c etc) the thread list is
> > re-enumerated, so new tasks are visible to GDB and tasks which have exited
> > are removed from the thread list.
> >
> > To achieve this GDB has to know about various Linux kernel data structures,
> > and also various fields within these structures (mainly task_struct,
> > thread_info, mm_struct). Code has been implemented to parse these structs,
> > and also do virtual to physical address translations and cache handling.
> > Various frame unwinders are also implemented to stop backtracing on various
> > exceptions, and entry points to the kernel (these would be a useful addition
> > regardless of which task awareness approach is taken).
> >
> > Advantages
> >  - Adds Linux kernel task awareness to GDB
> >  - Supports symbolic debugging of user processes
> >  - Contextual information (structs / field offsets) are readily available inside of GDB
> >  - Has been used and well tested inside ST for some time
> >
> > Disadvantages
> >  - Being implemented in C within GDB creates a dependency between GDB and the Linux kernel
> >  - Mainly tested on 3.10 and 2.6.30 with ARM and SH4 kernels, being upstreamed would
> >    expose it to many differing kernel versions and architectures.
> >
> >  - I can't see any other "OS awareness" support currently in the GDB code base
> 
> Actually, it seems that there's a bit of special OS kernel debug
> support, e.g. in bsd-kvm.c.

Thanks for the pointer I will take a look in there. It's useful to know where
all the OS specific bits in GDB are already.
> 
> But I'd rather view this as providing support for the Linux kernel
> *runtime* and compare it to components like linux-thread-db.c for
> user-space threads.  GDB has several of those already.  They often leave
> the internals of accessing the thread library's data structures to a
> "thread debug library" outside GDB.  We may or may not want to do
> something similar for the kernel, depending on the complexity of the
> data structures involved and on the likelihood for them to change.

Interesting..I wasn't aware of the use of an external library, I will take a
look in there to. Looking at the gdb-kdump implementation that someone kindly
CC'ed me in on the discussion it looks like that is doing something similar at
least for the virt to phys translation (which is something the LKD patches
currently do in GDB).

> Maybe the "thread debug library" could actually be a Python script
> provided by the kernel developers, as outlined in 1.3 below.  Or maybe
> it's just easier to put all the required logic into GDB and perform
> appropriate version checks where necessary.

It is certainly easier from an implementation PoV as we have access to all the
required objects inside GDB in C, which I don't think is the case currently with
the python interface. Also of course we have 2 existing out of tree
implementations (LKD and gdb-kdump). Having taken a brief look at the gdb-kdump
code that is also parsing the kernel structures in GDB C code.

> 
> > 1.2) OpenOCD / GDB Linux task awareness
> > =======================================
> >
> > Whilst looking for any prior art in this area I found that OpenOCD already
> > implements some basic Linux task awareness. See here: -
> >  http://openocd.org/doc/doxygen/html/linux_8c_source.html
> >
> > I've this working to the extent where I can connect via JTAG to a ARMv7 U8500
> > Snowball board and enumerate a thread list in GDB. It required some debugging
> > & hacking to get this far.
> >
> > This implementation passes the thread list to GDB via the gdbremote protocol
> > and as such changes required in GDB are minimal.
> >
> > Advantages
> >  - OpenOCD already supports many target types (ARM / MIPS), and has support
> > for virt to phys translations / cache handling etc.
> >
> >  - OpenOCD also implements task awareness for other RTOS’s (ThreadX / FreeRTOS / eCos)
> >
> >  - Using gdbremote means GDB changes are (so far) minimal.
> >
> >  - OpenOCD / Linux Kernel dependency already exists
> >
> >
> > Disadvantages
> >  - Creates a dependency between Linux kernel data structures and OpenOCD.
> >
> >  - I believe finding field offsets within structs is currently not possible via
> >    gdbremote protocol.
> >
> >    Currently OpenOCD generates these offsets at compile time which is ugly and needs
> >    fixing. See here http://openocd.org/doc-release/doxygen/linux__header_8h_source.html.
> >
> >    Being able to find field offsets would in my opinion be a useful addition to the
> >    gdbremote protocol which would allow OpenOCD task awareness to work much better
> >    at runtime.
> >    
> >  - Doesn't support debugging user processes. I think this would still require some
> >    GDB changes, and also gdbremote protocol changes to get working correctly.
> >
> >  - Needs to be made more generic
> >
> >
> > 1.3) Python Task awareness
> > ==========================
> >
> > Jan Kiszka from Siemens has implemented some basic Linux kernel task awareness using
> > the GDB Python interface.
> >
> > See here https://lwn.net/Articles/631167/ and here
> > https://github.com/torvalds/linux/blob/master/Documentation/gdb-kernel-debugging.txt
> >
> > This support is currently limited to the following commands: -
> >
> > lx_current -- Return current task
> > lx_per_cpu -- Return per-cpu variable
> > lx_task_by_pid -- Find Linux task by PID and return the task_struct variable
> > lx_thread_info -- Calculate Linux thread_info from task variable
> >
> > However this could be extended to build up a thread list. The GDB Python interface
> > would I believe need to be extended to allow a thread list to be passed back to GDB
> > via Python.
> >
> > Advantages
> >  - Code parsing Linux kernel structures lives in the kernel source tree
> >  - Contextual information is easily available
> >  - Works with all OCD implementations not just OpenOCD 
> >
> > Disadvantages
> >  - Doesn't exist yet
> >  - Python / GDB interface would need to be extended to support threads
> >
> > Questions
> > ========
> >
> > 1) What do GDB community think about having Linux OS task awareness in the GDB
> >    codebase?
> 
> For my part, I don't see problem with that.  We already have code in GDB
> for supporting various runtime environments.  This would just be another
> one.  Currently, even if an external thread debug library is involved,
> the code that uses it is still specially crafted for a particular
> runtime and does not work for another.

Ok, good to know that a C implementation inside GDB isn't necessarily a no-go
and that there are other implementations already merged. This route would
enable the most anmount of re-use with the existing LKD implementation.

Does anyone have experience / thoughts on how often the existing threading
implementations that are already inside GDB break? Presumably these are also
tightly coupled to parsing out of tree data structures (or the libraries which
GDB relies on to do this).

I'm trying to get a feel for what the current maintenance burden is like having
these implementations in C.

> 
> > 2) Is there a preferred way to implement this (C extension / gdbremote / Python
> >    / something else)?
> 
> Since many existing JTAG debuggers as well as Qemu will not offer a
> Linux task list, I think such logic belongs on the GDB client side.
> It's also on the client side where you have the DWARF info from vmlinux
> that enables dealing with varying field offsets in structures.
> 
> Whether the logic should be written in C or as a Python (or Guile)
> extension, I don't have a strong opinion.

Ok.

> 
> > 3) Any other implementations / advantages / disadvantages that I'm not currently
> >    aware of?
> >
> > 2) Loadable module support
> > ==========================
> >
> > 2.1) LKD Loadable module support
> > ================================
> >
> > lkd-modules.c adds better Linux kernel module symbolic symbol support to GDB using the
> > GDB shared libraries infrastructure (solib).
> >
> > It has hooks to enable the debugging of module __init sections to reflect that these pages
> > get discarded post-load. It also implements the same section layout algorithm as
> > the kernel to speed up symbol resolution, only inspecting the target’s memory if there is a
> > mismatch (inspecting target memory can be slow).
> >
> > I think this part could be upstreamed separately to the task awareness support, although
> > I’ve not tried separating it yet from the other LKD patches.
> 
> Yes, I think that separate patch sets would be helpful.
> 
> > Advantages
> >  - Allows full symbolic debugging of Linux loadable modules including init sections
> >
> >  - Would be independent of underlying communication mechanism (JTAG / gdbremote etc)
> >
> > Disadvantages
> >  - Some dependency between GDB and Linux kernel is still present
> >
> > Questions:
> >
> > Are GDB community happy to have a Linux specific solib functionality in the GDB code base?
> 
> For my part, same as with thread support: GDB already supports various
> shared library implementations and "knows" about their internals, so
> this would just be another one.

Ok good to know. I will take a read of the existing thread support already in GDB.
> 
> > 2.2) Python Loadable module support
> > ===================================
> >
> > I've not managed to look to much at this, but some basic support exists, see here
> > https://github.com/torvalds/linux/blob/master/Documentation/gdb-kernel-debugging.txt
> >
> > Having read the LKD modules implementation I don’t think this can be anywhere near
> > as functionally complete.
> >
> >
> > 3) Linux Helper commands
> > ========================
> >
> > 3.1) LKD helper commands
> > ========================
> >
> > LKD implements various Linux helper commands inside GDB such as: -
> >  - dmesg - dump dmesg log buffer from kernel
> >  - process_info - prints various info about current process
> >  - pmap - prints memory map of current process
> >  - vm_translate - translates virtual to physical address
> >  - proc_interrupts - prints interrupt statistics
> >  - proc_iomem - prints I/O mem map
> >  - proc_cmdline - prints the contents of /proc/cmdline
> >  - proc_version - prints the contents of /proc/version
> >  - proc_mounts - print the contents of /proc/mounts
> >  - proc_meminfo - print the contents of /proc/meminfo.
> >
> > Advantages
> >  - Can be used by all GDB based debug solutions
> >  - Some precedent of OS related commands in the GDB code base 
> >    https://sourceware.org/gdb/current/onlinedocs/gdb/OS-Information.html#OS-Information
> >
> > Disadvantages
> >  - Creates a GDB dependency with the kernel
> >
> > Python helper commands
> > ======================
> >
> > Jan’s python scripts also implement some of the same LKD commands such as 'dmesg'.
> >
> > See here https://lwn.net/Articles/631167/ and 
> > https://github.com/torvalds/linux/blob/master/Documentation/gdb-kernel-debugging.txt
> >
> > Advantages
> >  - Code that is parsing Linux kernel structures lives in the kernel source tree
> >  - Can be used by all GDB based debug solutions
> >
> >
> > Questions
> >  - Do GDB community mind Linux specific custom commands being added to GDB code base? 
> 
> I'm not sure about that one.  We also need to consider the capability
> for the user to add more such commands, e.g. specific to certain device
> drivers or certain debug scenarios.  In my view both cases should
> ideally be covered with the same underlying mechanism.
> 
> > My current opinion is that helper commands which can be, should be migrated from C code
> > into Python, and merged into the kernel source tree (and then retired from the LKD patchset).
> 
> That's my current view as well.  Or maybe we could consider using EPPIC
> for those.  (See my Cauldron slides.)

Ok, I'm glad we're aligned on that. Once again many thanks for your feedback.

regards,

Peter.