Hello Daniel,
Thanks very much for the write up, I think I understand it a lot better now :)
On Tue, Aug 5, 2025 at 8:54 PM Daniel P. Berrangé <berra...@redhat.com> wrote:
>
> On Tue, Aug 05, 2025 at 07:47:39PM +0300, Manos Pitsidianakis wrote:
> > On Tue, Aug 5, 2025 at 7:43 PM Daniel P. Berrangé <berra...@redhat.com>
> > wrote:
> > >
> > > On Tue, Aug 05, 2025 at 07:25:39PM +0300, Manos Pitsidianakis wrote:
> > > > On Tue, Aug 5, 2025 at 7:05 PM Daniel P. Berrangé <berra...@redhat.com>
> > > > wrote:
> > > > >
> > > > > On Mon, Aug 04, 2025 at 04:47:13PM +0300, Manos Pitsidianakis wrote:
> > > > > > This RFC series contains some simple patches I've been sitting on
> > > > > > for
> > > > > > some months to allow tracing in rust devices in a similar matter to
> > > > > > C,
> > > > > > only it's done via a proc-macro codegen instead of using tracetool
> > > > > > script or equivalent.
> > > > >
> > > > > IIUC, this series is only emitting the traces events via the
> > > > > qemu_log function, and so feels like it is missing the benefit
> > > > > of tracing, vs the traditional logging framework.
> > > > >
> > > > > In our RHEL & Fedora distro builds we disable the log backend
> > > > > and enable dtrace, so that we have fully dynamic tracing and
> > > > > observability across the kernel, qemu, libvirt and other
> > > > > components with dtrace integration.
> > > >
> > > > Hi Daniel,
> > > >
> > > > Thanks for the insight! Do you have any points where I should look at
> > > > the trace implementation for how the different backends are supported?
> > > >
> > > > So I think there's already work in progress to support proper tracing
> > > > for Rust, I only sent this as a temporary fixup to provide some kind
> > > > of parity between C and Rust implementations until a proper, better
> > > > solution is available that can replace it.
> > >
> > > Can the rust code not easily consume the existing functions in the
> > > trace.h files generated for the C code as a short-term solution ?
> > >
> > > It would not benefit from the code inlining in the same way as C
> > > would, but it would at least give feature parity for tracing with
> > > all the trace backends are available.
> > >
> > > Then, we can look at optimizing with a pure rust impl of some
> > > backends at a later date, to regain what we lost from lack of
> > > inlining ?
> >
> > It can, but we'd need to add extra intermediate steps to convert the
> > trace headers into Rust equivalent code, so it's not ideal.
> >
> > I tried to generate code exactly like the generated trace headers
> > though, so I'm not sure what is missing to be honest (hence my
> > previous email question). The generated code generates TraceEvents and
> > registers them with trace_event_register_group. What else is missing
> > to support e.g. dtrace?
>
> 'trace_event_register_group' is essentially irrelevant for the
> fully dynamic trace backends like dtrace - that's only used for
> the backends whose output is controlled by QEMU monitor commands
> / command line arguments.
>
> In the dtrace case the binary gets instructions which are a squence
> of nops normally, and dtrace tool gets the kernel to live patch the
> binary at runtime to put in a jump for any probes that are being
> watched.
>
> Take a look at the generated files <build-dir>/trace/trace-*.h when
> using the different '--enable-trace-backends=...' options.
>
> eg taking the trace-crypto.h header, with 'log' backend we see it
> emits
>
> if (trace_event_get_state(TRACE_QCRYPTO_TLS_SESSION_CHECK_CREDS) &&
> qemu_loglevel_mask(LOG_TRACE)) {
> #line 23 "../crypto/trace-events"
> qemu_log("qcrypto_tls_session_check_creds " "TLS session check creds
> session=%p status=%s" "\n", session, status);
> #line 372 "trace/trace-crypto.h"
> }
>
> but with dtrace it emits
>
> QEMU_QCRYPTO_TLS_SESSION_CHECK_CREDS(session, status);
>
> which is a referencing a macro created by the external 'dtrace' binary,
> which in the Linux case ends up looking like
>
> #if defined STAP_SDT_V1
> #define QEMU_QCRYPTO_TLS_SESSION_CHECK_CREDS_ENABLED() __builtin_expect
> (qcrypto_tls_session_check_creds_semaphore, 0)
> #define qemu_qcrypto_tls_session_check_creds_semaphore
> qcrypto_tls_session_check_creds_semaphore
> #else
> #define QEMU_QCRYPTO_TLS_SESSION_CHECK_CREDS_ENABLED() __builtin_expect
> (qemu_qcrypto_tls_session_check_creds_semaphore, 0)
> #endif
> __extension__ extern unsigned short
> qemu_qcrypto_tls_session_check_creds_semaphore __attribute__ ((unused))
> __attribute__ ((section (".probes")));
> #define QEMU_QCRYPTO_TLS_SESSION_CHECK_CREDS(arg1, arg2) \
> DTRACE_PROBE2 (qemu, qcrypto_tls_session_check_creds, arg1, arg2)
>
> you can end up enabling multiple trace backends concurrently too.
>
> If you're thinking this is all rather complicated, you'd be right,
> which is why for initial feature parity I figured the simplest is
> likely to just wrap the existing QEMU inline probe function, so
> Rust doesn't need to know about the different backends... yet...
Yes, that indeed makes sense. Generated C trace headers statically
linked to a standalone trace crate library for each subsystem, that
rust qemu crates can link to in return is the cleanest solution for
this approach IMHO, because doing this kind of codegen via macros
needs interaction with meson to generate the C sources and then run
bindgen all while compiling this one crate which is a single meson lib
target.
It might be possible to generate the equivalent of the C code for each
backend just like this RFC generates only the log backend code, I'll
take a look out of curiosity...
>
> FWIW, the original DTrace authors created a Rust crate with native
> rust integration of dynamic probes.
>
> https://github.com/oxidecomputer/usdt
>
> I think that (somehow) we probably want to integrate that with QEMU
> and its tracetool.
