Stefan Hajnoczi writes: > On Thu, Jul 27, 2017 at 11:40:17AM +0100, Peter Maydell wrote: >> On 27 July 2017 at 11:32, Stefan Hajnoczi <stefa...@redhat.com> wrote: >> > On Wed, Jul 26, 2017 at 03:44:39PM +0300, Lluís Vilanova wrote: >> >> And why exactly is this a threat? Because it can be used to "extend" QEMU >> >> without touching its sources? Is this a realistic threat? (it's a rather >> >> brittle >> >> way to do it, so I'm not sure it's practical) >> > >> > Unfortunately it is a problem. I recently came across a product that >> > was using LD_PRELOAD= to "integrate" with QEMU. People really abuse >> > these interfaces instead of integrating their features cleanly into >> > QEMU. >> >> ...if people who want to do this kind of thing already can and >> do use LD_PRELOAD for it, I don't think we should worry too much >> about trying to make the instrumentation plugin API bulletproof >> against similar abuse. >> >> > I see the use cases that Peter has been describing and am sure we can >> > come up with good solutions. What I care about is that it doesn't allow >> > loading a .so that connects to arbitrary trace events. >> >> That said, I agree that we don't really need an arbitrary-trace-event >> setup here, and we should probably design our API so that it isn't >> handing the trace plugin hooks pointers into QEMU's internals. >> We want an API that makes it easy for people to do things based on >> changes of the guest binary's state (registers, insns, etc etc) >> and which makes it hard for them to accidentally trip themselves up >> (eg by prodding around in QEMU internal data structures). >> This will have the secondary benefit that it's unlikely that future >> changes to QEMU will break plugin code. >> >> >> As a side note, I find instrumentation to be most useful for guest code >> >> events, >> >> which mostly contain non-pointer values (except for the CPUState*). >> >> For instance we definitely should not be passing a CPUState* to >> any plugin function.
> The gdbstub protocol has relevant features for accessing guest memory, > registers, etc. Perhaps a set of QEMU-specific events can be added > (e.g. tb generated) so it's possible to instrument and control the > guest from an instrumentation program (written in any language). > Perhaps there is a fundamental reason why this isn't possible due to the > protocol design, because using gdbstub halts all vcpus, etc. I don't > know. > Do you think this is an interesting direction? It definitely seems like > a powerful approach though performance would be less than running native > code inside the QEMU process. That's the same approach someone else dubbed as using a fifo with "synchronous" events, right? I have some measurements on this using a pipe, and overheads are 1000x to 2300x for each communication event (compared to a function call, and depending on whether each process/thread is pinned to the same or different CPU). Cheers, Lluis