在 2022/12/7 16:21, Marc Zyngier 写道:
On Tue, 06 Dec 2022 12:58:26 +0000,
chenxiang <chenxiang66@xxxxxxxxxxxxx> wrote:
From: Xiang Chen <chenxiang66@xxxxxxxxxxxxx>
Currently it only records statistical data for all vcpus, but we ofen want
to know statistical data for a single vcpu, there is no debugfs for that.
So add vcpu debugfs to record statstical data for every single vcpu, and
also enable vcpu debugfs for arm64.
After the change, those vcpu debugfs are as follows (we have 4 vcpu in the
vm):
[root@centos kvm]# cd 2025-14/
[root@centos 2025-14]# ls
blocking halt_wait_hist vcpu0
exits halt_wait_ns vcpu1
halt_attempted_poll halt_wakeup vcpu2
halt_poll_fail_hist hvc_exit_stat vcpu3
halt_poll_fail_ns mmio_exit_kernel vgic-state
halt_poll_invalid mmio_exit_user wfe_exit_stat
halt_poll_success_hist remote_tlb_flush wfi_exit_stat
halt_poll_success_ns remote_tlb_flush_requests
halt_successful_poll signal_exits
[root@centos 2025-14]# cat exits
124689
[root@centos 2025-14]# cat vcpu0/exits
52966
[root@centos 2025-14]# cat vcpu1/exits
21549
[root@centos 2025-14]# cat vcpu2/exits
43864
[root@centos 2025-14]# cat vcpu3/exits
6572
[root@centos 2025-14]# ls vcpu0
blocking halt_poll_invalid halt_wait_ns pid
exits halt_poll_success_hist halt_wakeup signal_exits
halt_attempted_poll halt_poll_success_ns hvc_exit_stat wfe_exit_stat
halt_poll_fail_hist halt_successful_poll mmio_exit_kernel wfi_exit_stat
halt_poll_fail_ns halt_wait_hist mmio_exit_user
This is yet another example of "KVM doesn't give me the stats I want,
so let's pile more stats on top". This affects every users (counters
are not free), and hardly benefits anyone.
Currently it already has vcpu debugfs on top, but it only records
statstical data for total vm
which is helpless for debug, for example, file exists records the number
of VM exist for all vcpus, before we encountered a
issue that there is something wrong with the thread of a vcpu which
doesn't VM exit but other vcpus are normal,
we can't get anything useful from current vcpu debugfs as the number of
exits still increase in current vcpu debugfs.
Compared with current vcpu debugfs, i think it is more useful to know
the statstical data for every vcpu and it benefits more.
How about you instead add trace hooks that allows you to plumb your
own counters using BPF or another kernel module? This is what is stuff
is for, and we really don't need to create more ABI around that. At
least, the other stat-hungry folks out there would also be able to get
their own stuff, and normal users wouldn't be affected by it.
Thanks,
M.
