[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-devel] Calculating real cpu usage of Xen domains correctly!
Hi John, John L Griffin wrote: My quick perusal of the current xeno-unstable code suggests that this SCHED_OP call is a null call. The SCHED_OP macro attempts to call the "do_block" function pointed to in the "struct scheduler sched_bvt_def" array, but this function pointer is never initialized, so it just does a NOP. Yes. I agree. It appears that what your patch does is limit when cpu_time gets updated, such that the time only gets updated when the exec_domain is in the BLOCKED state:No. Currently, cpu_time will always get updated for the prev domain when __enter_scheduler() is called. My patch modifies that behavior the behavior so that it only gets updated if prev->blocked is false or if prev->blocked is true but there are events pending for the domain. The last part may not be right. Sorry, this is my fault, my mail client badly munged the patch when I copy-pasted it. Let me show you the code:if ( test_bit(EDF_BLOCKED, &prev->ed_flags) ) prev->cpu_time += now - prev->lastschd; } if ( test_bit(EDF_BLOCKED, &prev->ed_flags) ) { /* This check is needed to avoid a race condition. */ if ( event_pending(prev) ) clear_bit(EDF_BLOCKED, &prev->ed_flags); else SCHED_OP(do_block, prev); } prev->cpu_time += now - prev->lastschd; Was changed to: if ( test_bit(EDF_BLOCKED, &prev->ed_flags) ) { /* This check is needed to avoid a race condition. */ if ( event_pending(prev) ) { clear_bit(EDF_BLOCKED, &prev->ed_flags); prev->cpu_time += now - prev->lastschd; } else SCHED_OP(do_block, prev); } else { prev->cpu_time += now - prev->lastschd; }It's not pretty at all. I'm not sure if the update after clear_bit() is necessary either. What this seems to be saying (in regard to your patch working) is that the cpu_time is updated when the domain relinquishes the CPU by block()ing, but cpu_time doesn't get updated if it relinquishes the CPU by yield()ing.No, it's the opposite. Sorry, I think the whitespace munging made the patch confusing. I wonder why this works. Is anyone familiar with block() vs yield(), that could lend some insight into what's going on?I'm quite confident this is the cause of the problem. I'm just not sure about that one cpu_time update. Regards, Anthony Liguori JLG ------------------------------------------------------- SF email is sponsored by - The IT Product Guide Read honest & candid reviews on hundreds of IT Products from real users. Discover which products truly live up to the hype. Start reading now. http://ads.osdn.com/?ad_id=6595&alloc_id=14396&op=click _______________________________________________ Xen-devel mailing list Xen-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.sourceforge.net/lists/listinfo/xen-devel
|
Lists.xenproject.org is hosted with RackSpace, monitoring our |