"Not correlated directly" ...
"1G-64G:128M,64G-1T:256M,1T-:512M"
Am I still asleep and dreaming? :)
Well, I said 'Not correlated directly', then gave sentences to explan
the reason. I would like to repeat them:
1) Crashkernel need more memory on some systems mainly because of
device driver. You can take a system, no matter how much memory you
increse or decrease total system RAM size, the crashkernel size needed
is invariable.
- The extreme case I have give about the i40e.
- And the more devices, narutally the more memory needed.
2) About "1G-64G:128M,64G-1T:256M,1T-:512M", I also said the different
value is because taking very low proprotion of extra memory to avoid
potential risk, it's cost effective. Here, add another 90M which is
0.13% of 64G, 0.0085% of 1TB.
Just let me clarify the problem I am having with all of this:
We model the crashkernel size as a function of the memory size. Yet,
it's pretty much independent of the memory size. That screams for "ugly".
The main problem is that early during boot we don't have a clue how much
crashkernel memory we may need. So what I see is that we are mostly
using a heuristic based on the memory size to come up with the right
answer how much devices we might have. That just feels very wrong.
I can understand the reasoning of "using a fraction of the memory size"
when booting up just to be on the safe side as we don't know", and that
motivation is much better than what I read so far. But then I wonder if
we cannot handle that any better? Because this feels very suboptimal to
me and I feel like there can be cases where the heuristic is just wrong.
As one example, can I add a whole bunch of devices to a 32GB VM and
break "crashkernel=auto"?
As another example, when I boot a 64G VM, the crashkernel size will be
512MB, although I really only might need 128MB. That's an effective
overhead of 0.5%. And especially when we take memory ballooning etc.
into account it can effectively be more than that.
Let's do a more detailed look. PPC64 in kernel-ark:
"2G-4G:384M,4G-16G:512M,16G-64G:1G,64G-128G:2G,128G-:4G";
Assume I would only need 385M on a simple 16GB VM. We would have an
overhead of ~4%. But maybe on ppc64 we do have to take the memory size
into account (my assumption and, thus, my comment regarding memory hotplug)?
I wonder if we could always try allocating larger granularity (falling
back to smaller if it fails), and once the kernel is able to come up
with a better answer how many devices there are and, thus, how big the
crashkernel area really should be, shrink the preallocated crashkernel
(either from the kernel or from user space)? Not completely trivial but
possible I think. It's trivial when we allocate a memmap for the
crashkernel (I think we mostly do but I might be wrong).
The "crashkernel=auto" would really do something magical good instead of
implement some heuristic base don the memory size.
[...]
So, all the rules we have are essentially broken because they rely
completely on the system RAM during boot.
How do you get this?
Crashkernel=auto is a default value. PC, VMs, normal workstation and server
which are the overall majority can work well with it. I can say the number
is 99%. Only very few high end workstation, servers which contain
many PCI devices need investigation to decide crashkernel size. A possible
manual setting and rebooting is needed for them. You call this
'essentially broken'? So you later suggestd constructing crashkernel value
in user space and rebooting is not broken? Even though it's the similar
thing? what is your logic behind your conclusion?
A kernel early during boot can only guess. A kernel late during boot
knows. Please correct me if I'm wrong.
Crashkernel=auto is mainly targetting most of systems, help people
w/o much knowledge of kdump implementation to use it for debugging.
I can say more about the benefit of crashkernel=auto. On Fedora, the
community distros sponsord by Redhat, the kexec/kdump is also maintained
by us. Fedora kernel is mainline kernel, so no crashkernel=auto
provided. We almost never get bug report from users, means almost nobody
use it. We hope Fedora users' usage can help test functionality of
component.
I know how helpful "crashkernel=auto" was so far, but I am also aware
that there was strong pushback in the past, and I remember for the
reasons I gave. IMHO we should refine that approach instead of trying to
push the same thing upstream every couple of years.
I ran into the "512MB crashkernel" on a 64G VM with memory ballooning
issue already but didn't report a BZ, because so far, I was under the
impression that more memory means more crashkernel. But you explained to
me that I was just running into a (for my use case) bad heuristic.
So system RAM size is the least important part to influence crashkernel
Aehm, not with fadump, no?
Fadump makes use of crashkernel reservation, but has different mechanism
to dumping. It needs a kernel config too if this patch is accepted, or
it can add it to command line from a user space program, I will talk
about that later. This depends on IBM's decision, I have added Hari to CC,
they will make the best choice after consideration.
I was looking at RHEL8, and there we have
fadump_cmdline = get_last_crashkernel(cmdline, "fadump=", NULL);
...
if (!fadump_cmdline || (strncmp(fadump_cmdline, "off", 3) == 0))
ck_cmdline = ...
else
ck_cmdline = ...
which was a runtime check for fadump.
Something that cannot be modeled properly at least with this patch here.
}
costing. Say my x1 laptop, even though I extended the RAM to 100TB, 160M
crashkernel is still enough. Just we would like to get a tiny extra part
to add to crashkernel if the total RAM is very large, that's the rule
for crashkernel=auto. As for VMs, given their very few devices, virtio
disk, NAT nic, etc, no matter how much memory is deployed and hot
added/removed, crashkernel size won't be influenced very much. My
personal understanding about it.
That's an interesting observation. But you're telling me that we end up
wasting memory for the crashkernel because "crashkernel=auto" which is
supposed to do something magical good automatically does something very
suboptimal? Oh my ... this is broken.
Long story short: crashkernel=auto is pure ugliness.
Very interesting. Your long story is clear to me, but your short story
confuses me a lot.
Let me try to sort out and understand. In your first reply, you asserted
"it's plain wrong when taking memory hotplug serious account as
we see it quite heavily in VMs", means you plain don't know if it's
wrong, but you say it's plain wrong. I answered you 'no, not at all'
with detailed explanation, means it's plain opposite to your assertion.
Yep, I might be partially wrong about memory hotplug thingy, mostly
because I had the RHEL8 rule for ppc64 (including fadump) in mind. For
dynamic resizing of VMs, the current rules for VMs can be very sub-optimal.
Let's relax "plain wrong" to "the heuristic can be very suboptimal
because it uses something mostly unrelated to come up with an answer".
And it's simply not plain wrong because in practice it gets the job
done. Mostly.
So then you quickly came to 'crashkernel=auto is pure ugliness'. If a
simple crashkernel=auto is added to cover 99% systems, and advanced
operation only need be done for the rest which is tiny proportion,
this is called pure ugliness, what's pure beauty? Here I say 99%, I
could be very conservative.
I don't like wasting memory just because we cannot come up with a better
heuristic. Yes, it somewhat gets the job done, but I call that ugly. My
humble opinion.
[...]
Yes, if you haven't seen our patch in fedora kexec-tools maining list,
your suggested approach is the exactly same thing we are doing, please
check below patch.
[PATCH v2] kdumpctl: Add kdumpctl estimate
https://lists.fedoraproject.org/archives/list/kexec@xxxxxxxxxxxxxxxxxxxxxxx/thread/YCEOJHQXKVEIVNB23M2TDAJGYVNP5MJZ/
We will provide a new feature in user space script, to let user check if
their current crashkernel size is good or not. If not, they can adjust
accordingly.
That's good, thanks for the pointer -- wasn't aware of that.
But, where's the current crashkernel size coming from? Surely
crashkernel=auto. You wouldn't add a random crashkernel size then
compared with the recommended crashkernel size, then reboot, will you?
If crashkernel=auto get the expected size, no need to reboot. Means 99%
of systems has no need to reboot. Only very few of systems, need reboot
after checking the recommended size.
Long story short. crashkernel=auto will give a default value, trying to
cover most of systems. (Very few high end server need check if it's
enough and adjust with the help of user space tools. Then reboot.)
Then we might really want to investigate into shrinking a possibly
larger allocation dynamically during boot.
Also: this approach here doesn't make any sense when you want to do
something dependent on other cmdline parameters. Take "fadump=on" vs
"fadump=off" as an example. You just cannot handle it properly as proposed
in this patch. To me the approach in this patch makes least sense TBH.
Why? We don't have this kind of judgement in kernel? Crashkernel=auto is
a generic mechanism, and has been added much earlier. Fadump was added
later by IBM for their need on ppc only, it relies on crashkernel
reservation but different mechanism of dumping. If it has different value
than kdump, a special hanlding is certainly needed. Who tell it has to be
'fadump=on'? They can check the value in user space program and add into
cmdline as you suggested, they can also make it into auto. The most suitable
is the best.
Take a look at the RHEL8 handling to see where my comment is coming from.
And I have several questions to ask, hope you can help answer:
1) Have you ever met crashkernel=auto broken on virt platform?
I have encountered it being very suboptimal. I call wasting hundreds of
MB problematic, especially when dynamically resizing of VMs (for
example, using memory ballooning)
Asking this because you are from Virt team, and crashkernel=auto has been
there in RHEL for many years, and we have been working with Virt team to
support dumping. We haven't seen any bug report or complaint about
crashkernel=auto from Virt.
I've had plenty of bug reports where people try inflating the balloon
fairly heavily but don't take the crashkernel size into account. The
bigger the crashkernel size, the bigger the issue when people try
squeezing the last couple of MB out of their VMs. I keep repeating to
them "with crashkernel=auto, you have to be careful about how much
memory might get set aside for the crashkernel and, therefore, reduces
your effective guest OS RAM size and reduces the maximum balloon size".
2) Adding crashkernel=auto, and the kdumpctl estimate as user space
program to get a recommended size, then reboot. Removing crashkernel=auto,
only the kdumpctl estimate to get a recommended size, always reboot.
In RHEL we will take the 1st option. Are you willing to take the 2nd one
for Virt platform since you think crashkernel=auto is plain wrong, pure
ugliness, essentially broken, least sense?
We are talking about upstreaming stuff here and I am wearing my upstream
hat here. I'm stating (just like people decades ago) that this might not
be the right approach for upstream, at least not as it stands.
And no, I don't have time to solve problems/implement solutions/upstream
patches to tackle fundamental issues that have been there for decades.
I'll be happy to help looking into dynamic shrinking of the crashkernel
size if that approach makes sense. We could even let user space trigger
that resizing -- without a reboot.
--
Thanks,
David / dhildenb
