I intend to change the NVMM API in order to reduce the data movements. The
patches are available here [1].
Basically until now the API expected the user to create structures and then
pass them in a libnvmm function. Typically:
/* The structures. */
struct nvmm_machine mach;
struct nvmm_x64_state state;
struct nvmm_event event;
struct nvmm_exit exit;
/* Create the machine and VCPU0. */
nvmm_machine_create(&mach);
nvmm_vcpu_create(&mach, 0);
/* Set RAX=123. */
nvmm_vcpu_getstate(&mach, 0, &state, NVMM_X64_STATE_GPRS);
state.gprs[NVMM_X64_GPR_RAX] = 123;
nvmm_vcpu_setstate(&mach, 0, &state, NVMM_X64_STATE_GPRS);
/* Inject an #UD exception. */
event.type = NVMM_EXCEPTION;
event.vector = 6;
nvmm_vcpu_inject(&mach, 0, &event);
/* Run. */
nvmm_vcpu_run(&mach, 0, &exit);
switch (exit.reason) {
...
}
This design has some problems: the structures can occupy a lot of stack
(nvmm_x64_state is > 1000 bytes), and must be copied in and out of the
comm page by libnvmm. We can avoid this by giving the user direct access
to the comm page.
A new nvmm_vcpu structure is introduced, and contains pointers to the comm
page. The new usage is:
/* The structures. */
struct nvmm_machine mach;
struct nvmm_vcpu vcpu;
/* Create the machine and VCPU0. */
nvmm_machine_create(&mach);
nvmm_vcpu_create(&mach, 0, &vcpu);
/* Set RAX=123. getstate populates vcpu.state->gprs[]. */
nvmm_vcpu_getstate(&mach, &vcpu, NVMM_X64_STATE_GPRS);
vcpu.state->gprs[NVMM_X64_GPR_RAX] = 123;
nvmm_vcpu_setstate(&mach, &vcpu, NVMM_X64_STATE_GPRS);
/* Inject an #UD exception. */
vcpu.event->type = NVMM_EXCEPTION;
vcpu.event->vector = 6;
nvmm_vcpu_inject(&mach, &vcpu);
/* Run. */
nvmm_vcpu_run(&mach, &vcpu);
switch (vcpu.exit->reason) {
...
}
No kernel changes are required. This should be part of NetBSD 9.
[1] https://m00nbsd.net/garbage/nvmm/nvmm-nocopy.zip
