Re: [PATCH 5/5] gdb: native target invalid architecture detection

[Andrew Burgess <aburgess@redhat.com>]

John Baldwin <jhb@FreeBSD.org> writes:

> On 5/31/22 7:30 AM, Andrew Burgess via Gdb-patches wrote:
>> If GDB is asked to start a new inferior, or attach to an existing
>> process, using a binary file for an architecture that does not match
>> the current native target, then, currently, GDB will assert.  Here's
>> an example session using current HEAD of master with GDB built for an
>> x86-64 GNU/Linux native target, the binary being used is a RISC-V ELF:
>> 
>>    $ ./gdb/gdb -q --data-directory ./gdb/data-directory/
>>    (gdb) file /tmp/hello.rv32imc.x
>>    Reading symbols from /tmp/hello.rv32imc.x...
>>    (gdb) start
>>    Temporary breakpoint 1 at 0x101b2: file hello.rv32.c, line 23.
>>    Starting program: /tmp/hello.rv32imc.x
>>    ../../src/gdb/gdbarch.h:166: internal-error: gdbarch_tdep: Assertion `dynamic_cast<TDepType *> (tdep) != nullptr' failed.
>>    A problem internal to GDB has been detected,
>>    further debugging may prove unreliable.
>> 
>> The same error is encountered if, instead of starting a new inferior,
>> the user tries to attach to an x86-64 process with a RISC-V binary set
>> as the current executable.
>> 
>> These errors are not specific to the x86-64/RISC-V pairing I'm using
>> here, any attempt to use a binary for one architecture with a native
>> target of a different architecture will result in a similar error.
>> 
>> Clearly, attempting to use this cross-architecture combination is a
>> user error, but I think GDB should do better than an assert; ideally a
>> nice error should be printed.
>> 
>> The problem we run into is that, when the user starts a new inferior,
>> or attaches to an inferior, the inferior stops.  At this point GDB
>> attempts to handle the stop, and this involves reading registers from
>> the inferior.
>> 
>> These register reads end up being done through the native target, so
>> in the example above, we end up in the amd64_supply_fxsave function.
>> However, these functions need a gdbarch.  The gdbarch is fetched from
>> the register set, which was constructed using the gdbarch from the
>> binary currently in use.  And so we end up in amd64_supply_fxsave
>> using a RISC-V gdbarch.
>> 
>> When we call:
>> 
>>    i386_gdbarch_tdep *tdep = gdbarch_tdep<i386_gdbarch_tdep> (gdbarch);
>> 
>> this will assert as the gdbarch_tdep data within the RISC-V gdbarch is
>> of the type riscv_gdbarch_tdep not i386_gdbarch_tdep.
>> 
>> The solution I propose in this commit is to add a new target_ops
>> method supports_architecture_p.  This method will return true if a
>> target can safely be used with a specific architecture, otherwise, the
>> method returns false.
>> 
>> I imagine that a result of true from this method doesn't guarantee
>> that GDB can start an inferior of a given architecture, it just means
>> that GDB will not crash if such an attempt is made.  A result of false
>> is a hard stop; attempting to use this target with this architecture
>> is not supported, and may cause GDB to crash.
>> 
>> This distinction is important I think for things like remote targets,
>> and possibly simulator targets.  We might imagine that GDB can ask a
>> remote (or simulator) to start with a particular executable, and the
>> target might still refuse for some reason.  But my thinking is that
>> these refusals should be well handled (i.e. GDB should give a user
>> friendly error), rather than crashing, as is the case with the native
>> targets.
>> 
>> For example, if I start gdbserver on an x86-64 machine like this:
>> 
>>    gdbserver --multi :54321
>> 
>> Then make use of this from a GDB session like this:
>> 
>>    $ ./gdb/gdb -q --data-directory ./gdb/data-directory/
>>    (gdb) file /tmp/hello.rv32imc.x
>>    Reading symbols from /tmp/hello.rv32imc.x...
>>    (gdb) target extended-remote :54321
>>    Remote debugging using :54321
>>    (gdb) run
>>    Starting program: /tmp/hello.rv32imc.x
>>    Running the default executable on the remote target failed; try "set remote exec-file"?
>>    (gdb)
>> 
>> Though the error is not very helpful in diagnosing the problem, we can
>> see that GDB has not crashed, but has given the user an error.
>> 
>> And so, the supports_architecture_p method is created to return true
>> by default, then I override this in inf_child_target, where I compare
>> the architecture in question with the default_bfd_arch.
>> 
>> Finally, I've added calls to supports_architecture_p for the
>> run (which covers run, start, starti) and attach commands.
>> 
>> You will notice a lack of tests for this change.  I'm not sure of a
>> good way that I can build a binary for a different architecture as
>> part of a test, but if anyone has any ideas then I'll be happy to add
>> a test here.
>
> Have you considered multi-arch cases such as running i386 binaries on an x86-64
> host or 32-bit arm binaries on an AArch64 host?  Will we need to override this
> method in certain targets (e.g. x86-linux-nat.c or x86-fbsd-nat.c) to support
> such cases?

For the x86 examples you gave, I think these all have the bfd_arch_i386
bfd architecture, so should work just fine.

But for the aarch64 case, I admit I don't know how this works.  A 32-bit
ARM binary is going to have bfd_arch_arm, while the AArch64 target will
be expecting a gdbarch with bfd_arch_aarch64.  But how GDB supports
running the former on the latter, I don't know.

I notice there's aarch64-linux-nat.c and aarch32-linux-nat.c, I wonder
if this has something to do with it...

Maybe someone with more ARM/AArch64 knowledge will chip in to fill in
some of the blanks.

But to answer the general question, if there is a case that the existing
code doesn't handle, then we can for sure override the new method in
specific *-nat.c targets in order to handle weird cases.

Thanks,
Andrew