I'm trying to wrap my head around how steal_account_process_tick() interacts
with account_process_tick().
Suppose we have CONFIG_VIRT_CPU_ACCOUNTING_GEN=y and CONFIG_NO_HZ_IDLE, with a
cpu hog on cpu0 to prevent it going idle.
As I understand it, account_process_tick() will be called once per tick to
decide whether that tick should be allocated against user/system/idle. However,
it first calls steal_account_process_tick() and then returns if that returns a
nonzero value.
The thing is, steal_account_process_tick() returns units of cputime, which I
think is nanoseconds on x86_64. So if we have a tiny amount of stolen time it
seems like that will prevent a whole tick from being accounted into
user/system/idle.
I feel like I must be missing something here, can someone tell me what it is?
Chris
On 03/04/2016 01:51 PM, Chris Friesen wrote:
> The thing is, steal_account_process_tick() returns units of cputime, which I
> think is nanoseconds on x86_64. So if we have a tiny amount of stolen time it
> seems like that will prevent a whole tick from being accounted into
> user/system/idle.
>
> I feel like I must be missing something here, can someone tell me what it is?
Looking at commit dee08a72 (from 2014) it seems like the units of the return
value of steal_account_process_tick() changed from ticks to cputime_t. I don't
see an equivalent change in the logic in account_process_tick(), which seems to
assume that a nonzero return value in steal_account_process_tick() means a whole
tick has been stolen.
Was there a change to make paravirt_steal_clock() increment in ticks? If not it
seems like there's a unit mismatch here.
Chris