[Crash-utility] [PATCH] Fix machdep->HZ calculation for kernel versions > 2.6.0

[HAGIO KAZUHITO(萩尾 一仁)]

-----Original Message-----
> 在 2021年01月12日 16:24, HAGIO KAZUHITO(萩尾　一仁) 写道:
> > Hi Bhupesh,
> >
> > -----Original Message-----
> >> We have hard-coded the HZ value for some ARCHs to either 1000 or 100
> >> (mainly for kernel versions > 2.6.0), which causes 'help -m' to show
> >> an incorrect hz value for various architectures.
> >
> > Good catch.  but seems crash uses (cfq_slice_async * 25) for machdep->hz
> > if it exists (please see task_init()), RHEL7 has it, but RHEL8 does not.
> > What do you see on RHEL8 for x86_64 with your patch?
> >
> 
> The symbol 'cfq_slice_async' has been removed from upstream kernel:
> f382fb0bcef4 ("block: remove legacy IO schedulers")
> 
> And RHEL8 also removed it.
> 
> > We should search for an alternate way like the current one first.
> >
> 
> Currently, there are several ways to get the value of HZ as below:
> 
> [1] calculate hz via the symbol 'cfq_slice_async'
>     But this symbol has been removed from upstream kernel

According to [0] below, the 'cfq_slice_async' cannot be used for the HZ
calculation on 4.8 and later kernels.  I've not found a perfect alternate,
but how about using 'bfq_timeout' for 4.12 and later including RHEL8?

const int bfq_timeout = HZ / 8;

     RELEASE: 4.18.0-80.el8.x86_64

crash> p bfq_timeout
bfq_timeout = $1 = 125

This value has not been changed since its introduction (aee69d78dec0).
Recent kernels configured with CONFIG_IOSCHED_BFQ=y can be covered with this?

[0] https://listman.redhat.com/archives/crash-utility/2021-April/msg00026.html

Thanks,
Kazu


> 
> [2] hardcode hz with the value 1000 (if kernel version > 2.6.0)
> 
> [3] get the hz value from vmcore, but that relies on kernel config
>     such as CONFIG_IKCONFIG, etc.
> 
> [4] Use sysconf(_SC_CLK_TCK) on some arches, not all arches.
>     See the micro definition of HZ in the defs.h
> 
> There seems to be no perfect solution. Any ideas?
> 
> 
> Thanks.
> Lianbo
> 
> > Thanks,
> > Kazu
> >
> >>
> >> I tested this on ppc64le and x86_64 and the hz value reported is 1000,
> >> whereas the kernel CONFIG_HZ_100 is set to Y. See some logs below:
> >>
> >> crash> help -m
> >>               flags: 124000f5
> >>
> (KSYMS_START|MACHDEP_BT_TEXT|VM_4_LEVEL|VMEMMAP|VMEMMAP_AWARE|PHYS_ENTRY_L4|SWAP_ENTRY_L4|RADIX_MMU|OP
> >> AL_FW)
> >>              kvbase: c000000000000000
> >>   identity_map_base: c000000000000000
> >>            pagesize: 65536
> >>           pageshift: 16
> >>            pagemask: ffffffffffff0000
> >>          pageoffset: ffff
> >>           stacksize: 16384
> >>                  hz: 1000
> >>                 mhz: 2800
> >>
> >> [host at rhel7]$ grep CONFIG_HZ_100= redhat/configs/kernel-3.10.0-ppc64le.config
> >> CONFIG_HZ_100=y
> >>
> >> Fix the same by using the sysconf(_SC_CLK_TCK) value instead of the
> >> hardcoded HZ values depending on kernel versions.
> >>
> >