On Sun, 30 Apr 2023 at 20:45, Francesco Cagnin <francesco.cag...@gmail.com> wrote: > > > Other than that, the structure of the patch looks OK, but > > I think you need to identify the cause of the problems > > with SMP setups that you mention in the cover letter, > > since they suggest that there's a bug lurking in here > > somewhere. > > In the current patch, in `hvf_arch_update_guest_debug()` I'm enabling > exiting the guest on debug exceptions only for the vCPUs that have inserted > software/hardware breakpoint or are singlestepping. In SMP setups this logic > looks flawed, since for example if vCPU #1 sets a software breakpoint and > vCPU #2 hits it, the generated debug exception for vCPU #2 will not exit the > guest and lead to panic for unexpected BRK. A possible fix is enabling > exiting the guest on debug exceptions for all vCPUs (and not just the ones > that have inserted breakpoints)—is this the way to go?
Yes, I think that makes sense. We treat breakpoints as VM-wide (and writing a brk instruction into guest memory is by definition VM-wide), so we should put all the vCPUs into "we're debugging the VM mode", not just one. > There's also a second analogous issue for which it feels like I'm missing > something. If through GDB a software breakpoint is inserted from vCPU #1 and > later vCPU #2 hits it, then when trying to resume execution after the hit > GDB fails with 'Cannot remove breakpoints', due to > `hvf_find_sw_breakpoint()` returning error because it (correctly) doesn't > find any software breakpoint for vCPU #2 (queue > `cpu->hvf->hvf_sw_breakpoints`). A possible fix seems to be modifying > `hvf_find_sw_breakpoint()` so that it searches for the breakpoint on all > vCPUs' queues; but I've skimmed through the analogous routines for TCG and > KVM and there's nothing resembling this fix, so I wonder why TCG and KVM > don't fail on my example GDB scenario? For KVM the sw breakpoint list is in cpu->kvm_state, and we share a single kvm_state across every vCPU in the VM (kvm_init_vcpu() initializes cpu->kvm_state to be a pointer to the singleton global kvm_state). Isn't cpu->hvf shared between vCPUs in the same way? For TCG breakpoints are effectively always hw breakpoints, regardless of whether gdb asks for a sw bp or a hw bp; the code path is different from how KVM and HVF do things. thanks -- PMM
