other: 0.171
semantic: 0.158
performance: 0.112
device: 0.107
vnc: 0.065
files: 0.055
graphic: 0.055
boot: 0.055
debug: 0.051
PID: 0.046
socket: 0.044
network: 0.031
permissions: 0.027
KVM: 0.023
performance: 0.215
boot: 0.158
debug: 0.114
files: 0.110
semantic: 0.099
other: 0.089
device: 0.052
PID: 0.051
network: 0.025
vnc: 0.022
KVM: 0.019
socket: 0.017
graphic: 0.015
permissions: 0.013

Running with -rtc clock=vm,base=<datetime> introduces arbitrary base shift at guest startup

When specifying 'base' for RTC to start with, it has incorrect implementation in combination with clock=vm.

I inspected source code. This is because it uses host clock (qemu_time() function return value) as reference with 'rtc_date_offset' operations across several places in code before guest execution starts, which has no relevance with clock=vm.

It works in vast majority of cases only thanks to combination of some luck and fast execution of qemu initialization phase relative to host real time (i.e. multiple calls of qemu_time() returns same value). But if qemu execution is being slow down by high CPU load on host or started just before value of second changes, it may accumulate at least 1 second (and in hard circumstances even 2+ seconds) of delay from specified base datetime before guest becomes ready to start.

This behavior breaks determinism of guest execution in icount mode. (Even if guest doesn't cares about precise time, just one second shift may trigger such chain of changes which accumulates significant difference in guest state at the moment when guest OS finishes booting.)

Why I didn't posted patch to qemu-devel ?
I have no idea how to patch it correctly, because it isn't clear how these things are expected to work when referenced across all qemu code and different use cases. Should vl.c/qemu_get_timedate() just be fixed ? Does each caller expect same behavior from it? Or only selected hw/* RTC implementations (such as hw/timer/mc146818rtc.c) have to be modified to use alternative function ? Or maybe it isn't specific to clock=vm and should be considered bad behavior in any case (i.e. time shouldn't advance before guest execution being started)?

Fix has been included here:
https://git.qemu.org/?p=qemu.git;a=commitdiff;h=eb6a52099160f1a6e66d7e