This patch introduce a new mode of IO request completion, polling mode.
Now, the sketch of ioreq process can be,
A. UOS vcpu0 generate PIO/MMIO ->
B. pcpu1(vcpu0 of UOS) trap into HV ->
C. pcpu1 build ioreq, send IPI and enter idle ->
D.1 pcpu0(vcpu0 of SOS) response IPI,
D.2 pcpu0 handle the ioreq in HV, kernel, DM,
D.3 pcpu0 mark ioreq as complete,
D.4 pcpu0 hypercall to enter HV ->
E.1 pcpu0 send IPI to wake pcpu1 up
E.2 UOS vcpu0 continue running
With this change, it skips D.4, E.1 steps. In step C, pcpu1 will enter a
polling ioreq state idle after send out the IPI.
It can save about ~5000 cpu cycles.
In polling mode, we do the polling in idle instead of pause cpu all the
time. It will consume more power. A better way is to use monitor/mwait
instructions which can put cpu into a sleep state with monitoring a
memory address. Unfortunately, APL has bug with monitor. We can gather
all ioreqs state into one monitorable memory and take advantage of
monitor/mwait for future platform.
The way polling or notification is per VM. We can config VMs in
different mode. By default, IO request completion will use notification
mode for all VMs. We can switch it by Kconfig.
Tracked-On: #1821
Signed-off-by: Shuo Liu <shuo.a.liu@intel.com>
Reviewed-by: Eddie Dong <eddie.dong@intel.com>
Acked-by: Anthony Xu <anthony.xu@intel.com>
e8296dce05