3:18 a.m.
On Thu, May 31, 2018 at 11:19:04AM -0400, John Ferlan wrote:
On 05/31/2018 08:14 AM, Martin Kletzander wrote:
> On Wed, May 30, 2018 at 12:00:26PM -0400, John Ferlan wrote:
>>
>>
>> On 05/21/2018 11:00 AM, Martin Kletzander wrote:
>>> TSEG (Top of Memory Segment) is one of many regions that SMM (System
>>> Management
>>> Mode) can occupy. This one, however is special, because a) most of
>>> the SMM code
>>> lives in TSEG nowadays and b) QEMU just (well, some time ago) added
>>> support for
>>> so called 'extended' TSEG. The difference to the TSEG implemented
in
>>> real q35's
>>> MCH (Memory Controller Hub) is that it can have any size from 1 MiB
>>> up to 65534
>>> MiB in 1 MiB increments. But more about that in the QEMU patch.
>>
>> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
>>
>> Which some reader, but not this one, may be eager to find ;-)
>>
>> Still is there a valid range QEMU would accept? Or do we just let QEMU
>> fail if the range is too high?
>>
>> I think QEMU has MCH_HOST_BRIDGE_EXT_TSEG_MBYTES_MAX
>>
>
> Rather than promising some value, I adjusted it so that it is no longer
> false,
> no matter what the max is there.
>
>>>
>>> Signed-off-by: Martin Kletzander <mkletzan(a)redhat.com>
>>> ---
>>> docs/formatdomain.html.in | 39 +++++++++++++++++++
>>> docs/schemas/domaincommon.rng | 5 +++
>>> src/conf/domain_conf.c | 60 ++++++++++++++++++++++++++++-
>>> src/conf/domain_conf.h | 1 +
>>> tests/genericxml2xmlindata/tseg.xml | 23 +++++++++++
>>> tests/genericxml2xmltest.c | 2 +
>>> 6 files changed, 129 insertions(+), 1 deletion(-)
>>> create mode 100644 tests/genericxml2xmlindata/tseg.xml
>>>
>>
>> In the category of I hate it when that happens, git am -3 "merged" in
>> two chunks incorrectly! Probably wouldn't have happened if I'd done
>
> You can enable/disable 3-way merges if you do (not) like them.
>
>> this sooner! The virDomainDefFeaturesCheckABIStability hunk got merged
>> into virDomainRedirFilterDefCheckABIStability and the tseg grammar got
>> merged under "vmport" and just before "gic". As you can
imagine the
>> results weren't pretty ;-).
>>
>
> Yeah, happened to me as well, I should've resent this, but I forgot
> about the
> merge issue and I also wanted to show that this was posted way before the
> freeze. Anyway, it's pointless to force it now, I'll leave it for later
> (meaning "after the release").
>
*Lots* of things were posted after this that got reviewed and pushed
before this - that just seems to be "the norm" (for whatever reason)...
Not receiving timely reviews is one of those areas that really bothers
me. I've been making the effort more recently to try to go in order of
longest without review - sometimes a ping makes me lose order though.
> Anyway, I keep my branches updated (every now and then) on my github
> repo [1],
> so if you want to check that, you always can.
>
> [1] https://github.com/nertpinx/libvirt
>
OK - I've added that as a remote...
>>> diff --git a/docs/formatdomain.html.in b/docs/formatdomain.html.in
>>> index 403b638bd4bd..39ebfe398bd7 100644
>>> --- a/docs/formatdomain.html.in
>>> +++ b/docs/formatdomain.html.in
>>> @@ -1897,6 +1897,9 @@
>>> <ioapic driver='qemu'/>
>>> <hpt resizing='required'/>
>>> <vmcoreinfo state='on'/>
>>> + <smm state='on'>
>>> + <tseg unit='MiB'>48</tseg>
>>> + </smm>
>>> </features>
>>> ...</pre>
>>>
>>> @@ -2056,6 +2059,42 @@
>>> <code>off</code>, default <code>on</code>)
enable or disable
>>> System Management Mode.
>>> <span class="since">Since 2.1.0</span>
>>> +
>>> + Optional sub-element <code>tseg</code> can be used to
>>> specify the
>>> + amount of memory dedicated to SMM TSEG. The size can be
>>> specified as a
>>> + value of that element, optional attribute
>>> <code>unit</code> can be
>>> + used to specify the unit of the aforementioned value
>>> (defaults to
>>> + 'MiB').
>>> +
>>
>> If this is to be a true paragraph break then these paragraphs needs to
>> be wrapped in <p> ... </p>; otherwise, this just becomes one long run
on
>> (and quite ugly IMO) paragraph.
>>
>>> + This value is configurable due to the fact that the
>>> calculation cannot
>>> + be done right with the guarantee that it will work
>>> correctly. For
>>> + QEMU TSEG was disabled up to and including
>>> <code>pc-q35-2.9</code> (it
>>> + does not make sense fo any other machine type than q35).
>>
>> s/fo/for/
>>
>> This also appears to be a paragraph break...
>>
>>> + From <code>pc-q35-2.10</code> the default value was
>>> changed to 16 MiB.
>>
>> s/From/Starting with/ ??? (not required, just a though)
>>
>>> + That should suffice for up to 272 VCPUs, 5 GiB guest RAM
>>> in total, no
>>> + hotplug memory range, and 32 GiB of 64-bit PCI MMIO
>>> aperture. Or for
>>> + 48 VCPUs, with 1TB of guest RAM, no hotplug DIMM range,
>>> and 32GB of
>>> + 64-bit PCI MMIO aperture. The values may also vary based
>>> on the loader
>>> + the VM is using.
>>> +
>>> + Additional size might be needed for significantly higher
>>> VCPU counts
>>> + or increased address space (that can be memory, maxMemory,
>>> 64-bit PCI
>>> + MMIO aperture size; roughly 8 MiB of TSEG per 1 TiB of
>>> address space)
>>> + which can also be rounded up.
>>
>> Uh, oh, hmmm... We seem to have this (perhaps more recent) "thing"
about
>> libvirt having to supply some attribute based on some (mostly difficult
>> to describe) algorithm that QEMU would use in order to create the
>> "optimum" size or use for some alternate algorithm. Of course, a few
>> libvir-list patches like that have been NACK'd because of the feeling
>> that providing a useful algorithm for a user to "decide upon" a
>> satisfactory attribute value cannot really be done. Off the top of my
>> head I can come up with two:
>>
>
> It's kind of a different story. Think of this as a memory size. You
> cannot determine the "right" amount of memory the VM should have. You
> can try to boot with X and double it until the OS installation succeeds.
> And hope you won't need to change it later.
>
>> 1. Add poll-max-ns property of each iothread:
>> https://www.redhat.com/archives/libvir-list/2017-February/msg01047.html
>>
>
> This is about tunables. It might change the performance/latency of QEMU
> slightly, but that's about it.
>
and there are those that would find it useful to have (bz 1545732)... If
Much of what I read (underscore emphasis mine) suggests otherwise:
- "There is a lot of sentiment *against* providing too many low level knobs
like this _without proper guidance_ on how they should be set."
- "To address this issue QEMU implements self-tuning algorithm that modifies
the current polling time to _adapt to different workloads_ and it can also
fallback to blocking syscalls."
- "The QEMU commits say the tunables all default to sane parameters so I'm
inclined to say we ignore them at the libvirt level entirely."
- "I'm fine if libvirt doesn't add a dedicated API for setting
<iothread>
polling parameters. It's _unlikely_ that users will need to change the
setting. In an emergency (e.g. disabling it due to a performance
regression) _they can_ use <qemu:arg value='-newarg'/>."
The only points for the polling to be enabled were along the lines of:
It _may_ help in _some_ workloads when you want a bit more throughput
for the price of more CPU cycles.
With vague definitions of how much CPU, throughput and without
description of how to find out if a particular workload fits this. Even
when all of that is there, then you need yet another explanation on how
to calculate the value to be set. And then it all goes down back to the
fact that QEMU is already doing some automated balancing for this
(because they can, because this is not part of the guest ABI). That way
you can never actually say if it will help and how much.
So for this one it is a clear "NO".
you don't have enough memory and your VM is paging like crazy, you
just
add more memory. Requires a reboot. Likewise, if your VM doesn't boot
you add/alter the magic TSEG value using some algorithm as described
above. From a 90,000 foot customer view is there a difference? It's just
a knob that the hypervisor has to allow something to be accomplished for
which libvirt provides the attribute to fine tune.
Yeah, you're right. That's why I think both of them should be exposed. Some
small differences to other knobs, just for completeness:
- by the time you realize that the VM doesn't have enough memory, it might be
too late as reboot isn't that easy of a thing for some production workloads
- on the other hand, you have a way to see that happening (compare it to the
polling interval above which you have no idea without proper benchmarks)
One more thing that's common to the memory size (and I hope TSEG in the future)
is that in mgmt apps the TSEG setting already has a place where to live and it
is exactly where the memory size lives currently. In templates. You have
"small vm" teplate and "ginormous vm". For the latter one you can
just add a
setting of TSEG _once_ per file. How's TSEG better and easier than memory? You
figure it once for the VM settings (and possibly firmware, but that's not going
to change much) and then it doesn't depend on the workflow, not even a little
bit.
Anyway.
What I see as the differences between tunables that make it in and tunables that
don't is that:
- the former are usually understandable and easy to see what they are in bare
metal. Everyone knows what memory is in the hardware, how it looks like, how
much is "not enough" and how much is "more that needed". We are
used to
those things back from the hardware times, even to changing them.
- The latter is usually something we were not able control in HW or didn't even
know it existed. For virtual workloads it might be completely different, but
sometimes people are forgetting that.
>> 2. Add support for qcow2 cache (many times, but most
recently):
>> https://www.redhat.com/archives/libvir-list/2017-September/msg00553.html
>>
>
> Similarly here, it allows setting something that can be (at least
> slightly) abstracted and in the worst case the performance will be
> slightly hindered.
This one I understand more why it would be rejected, but still providing
the value allows certain things to work a whole lot better. I also know
Berto has been "fine tuning" the algorithm in later QEMU releases - so
that's like hitting a moving target.
This is very similar, it's just that there is no automatic balancing
done by QEMU. But it usually is also about how you write the docs. The
option can make very much sense, but if someone writes "Setting asdf can
allows fine-tuning of the asdf value in the underlying hypervisor", then
no matter how much that value makes sense it is not reflected in the
docs. That's why I tried to add all the relevant info into the docs so
that it's clear what it is doing, how to set it, to what values and
when.
Apart from the fact that there is a "link" to some file in the QEMU
repository that someone is supposed to read, plus the decision for the
value determination are written there (but not why they are not
automatically calculated, or maybe I missed it), it:
- is not possible to try using the <qemu:arg value='-newarg'/> approach
- the docs say:
<b>In general you should leave this option alone, unless you
are very certain you know what you are doing.</b>
So in this particular case I wouldn't be totally against having it
there. If you don't want to use it, then "just don't touch that" is an
approach that shouldn't hurt anyone.
>
> Contrast this with TSEG which, in case it is set incorrectly, will
> prevent the machine from booting at all. If we go to the extreme, not
> only can you easily try to find out the right amount to set for a
> particular machine, but you can even do that programmatically since when
> OVMF fails due to small extended TSEG size it will reboot very fast.
> And you can get that in form of events. When I tried it now it even
> looks like you get rtc-change event when the domain doesn't reboot
> immediately due to small TSEG size.
Is there something that is provided that would give the "hint" to adjust
the TSEG size?
I thought there was, but maybe I'll add some more info after the discussion.
>
> I will not put it in the docs because I will not guarantee that this is
> the right way to go, but this is how events look for default TSEG size
> for a guest that needs a lot more (it has 240 possible vCPUs and 256 TiB
> of maximum memory, but because it starts with only 1 vCPU and 1GiB
> memory I can try it out easily on my machine:
>
> virsh # event --domain nixos --all --loop --timestamp
> 2018-05-31 09:42:21.060+0000: event 'lifecycle' for domain nixos:
> Resumed Unpaused
> 2018-05-31 09:42:21.066+0000: event 'lifecycle' for domain nixos:
> Started Booted
> 2018-05-31 09:42:21.514+0000: event 'reboot' for domain nixos
> 2018-05-31 09:42:21.964+0000: event 'reboot' for domain nixos
> 2018-05-31 09:42:22.414+0000: event 'reboot' for domain nixos
> 2018-05-31 09:42:22.868+0000: event 'reboot' for domain nixos
> 2018-05-31 09:42:23.325+0000: event 'reboot' for domain nixos
> 2018-05-31 09:42:23.778+0000: event 'reboot' for domain nixos
> 2018-05-31 09:42:24.230+0000: event 'reboot' for domain nixos
> 2018-05-31 09:42:24.681+0000: event 'reboot' for domain nixos
> ...
>
> you get the point.
>
> And this is how it looks when I start it with increased size:
>
> virsh # event --domain nixos --all --loop --timestamp
> 2018-05-31 09:43:24.578+0000: event 'lifecycle' for domain nixos:
> Resumed Unpaused
> 2018-05-31 09:43:24.584+0000: event 'lifecycle' for domain nixos:
> Started Booted
> 2018-05-31 09:43:31.808+0000: event 'rtc-change' for domain nixos: 0
> 2018-05-31 09:43:32.808+0000: event 'rtc-change' for domain nixos: 0
>
> The reasons for this not being done automatically are (from the top of
> my head):
>
> - The above is just something I figured out myself, but it's not the
> recommended way written anywhere. Maybe I'm wrong and it doesn't
> really work, but it still can be done manually.
>
> - You cannot change it however you would like automatically, it is part
> of the guest ABI and we are striving for keeping that stable.
>
> - Trying to figure this out by 1 MiB increments might take some time,
> but increasing it faster might be wasteful.
>
> Basically there is no one-size-fits-all value, no easy way to do it
> automatically (maybe what I tried), but very good explanation how to do
> that manually and very easy way to do that. Also, from the SW POV, it
> doesn't even depend on the guest OS, just on the loader/bios so if you
> have two same domains (like a template in OpenStack for example) you try
> it out once and then you have the value that just works and will
> continue working until you change something for the domain. And what
> it depends on is clearly written in the documentation.
>
And there are those that could say if the underlying hypervisor knows
that for certain memory sizes and/or vCPU counts that the TSEG will be
too small for specific machine types that then the underlying hypervisor
should be the one to "choose" a value that's programatically appropriate
which to a degree IIUC is the argument being used against allowing a
libvirt knob for the poll-max-ns and qcow2 cache sizes.
And they would be wrong as for TSEG the hypervisor a) doesn't know that
and b) cannot change that once it was started.
>>> +
>>> + Due to the nature of this setting being similar to "how
>>> much RAM
>>> + should the guest have" users are advised to either consult
>>> the
>>> + documentation of the guest OS or loader (if there is any),
>>> or test
>>> + this by trial-and-error changing the value until the VM boots
>>> + successfully. Yet another guiding value for users might
>>> be the fact
>>> + that 48 MiB should be enough for pretty large guests (240
>>> VCPUs and
>>> + 4TB guest RAM), but it is on purpose not set as default as
>>> 48 MiB of
>>> + unavailable RAM might be too much for small guests (e.g.
>>> with 512 MiB
>>> + of RAM).
>>
>> and this is the exact reason why patches like this get NACKd - because
>> trial and error should not be a 'desired' means to calculate.
>
> It is not. They are rejected because either a) there is no
> documentation how to properly check if the value is the right value when
> doing trial-end-error (this is not the case here since you can see if
> the machine boots or not) or b) the values being set are too specific
> instead of being abstracted -- setting value in KiB between 0 and the
> size of a disk instead of "max_performance" or "min_latency"
(this is
> not the case here, the documentation explains what the size is and why
> it is not about few guessable values).
>
> Basically we are NACKing simple pass-through values without
> understanding them and adding some documentation for them. For example
> stuff for which we have documentation along the lines of: "Element asdf
> can be used to set the asdf of the domain."
>
I guess it's all a matter of how things are being viewed. I can
certainly see how someone without a redhat.com email may view this
series as similar to the referenced ones. It's more about mitigating
failure possibilities by being able to provide knobs that can help (or
get you into trouble). Not everyone has large configurations - so do
they really need the value? Not everyone needs every knob, but having
them available for when you do need them isn't a bad thing.
>> bz referenced in patch 5 has an incredible amount of data and
>> calculations that provide even more insight and details that are lost
>> when we try to summarize in a libvirt meaningful patch.
>>
>
> Let me know what relevant information from the bz you are missing in the
> documentation and I'll gladly add it.
>
If you think what's provided above is enough, then leave it as is. If
there's something in that bz or in the ensuing discussion, then add it.
Even though it's difficult when you're in the middle of the code - if
you were someone who ran into this situation without knowing anything
about it - would what's provided above give enough information to help
you decide which value to use? That whole trial and error comment is
what leaves me wondering as a reader who knew nothing about this before
I started. And now I have more I need a pub to help forget ;-)
Here comes the weekend!
>> What it seems is really needed is an attribute that libvirt
provides
>> that tells QEMU to calculate the optimum size.
>>
>>> +
>>> + See <a href="#elementsMemoryAllocation">Memory
Allocation</a>
>>> + for more details about the <code>unit</code>
attribute.
>>> + <span class="since">Since 4.5.0</span> (QEMU
only)
>>
>> haha - see you put 4.5.0 and this is the 4.4.0 release - so it was
>> ignored until 4.5.0 was "on the clock" ;-)
>>
>> Ironically this is pointed out as QEMU only; however, genericxml2xmltest
>> is used/updated.
>>
>
> So?
>
Why would someone add this generically if it only works for QEMU?
The part where the VM is set up based on that is only supported in QEMU.
The part that does formatting/parsing is done in src/conf/domain_conf.c,
which is used by all drivers.
I would add it there because it:
a) might be later added to other hypervisor drivers
b) even when someone is compiling with the qemu driver turned off they
cannot break it (provided they do run tests)
>> So, I personally don't mind if this attribute is added;
however, I think
>> we become hypocrites to a certain degree if patches continue to be
>> blocked/NACKed to other subsystems that have attributes that allow
>> certain amount of control, but don't come with exact sizing references.
>> Still if this is pushed, then perhaps those others can use this as the
>> means to provide a reference to other knobs added.
>>
>
> How much more exact would you wanted to be in terms of sizing? If it
> could be any more exact we wouldn't need the tunable at all. Please
> don't compare it to other tunables that we didn't want exposed just
> because it "sounds similar". I hear lot of people just put stuff like
> this into "unknown knobs" box and they treat it the same. But there are
> differences and it's not all
>
>> You can have my :
>>
>> Reviewed-by: John Ferlan <jferlan(a)redhat.com>
>>
>
> I wasn't ever against the R-b, but I guess I'm missing the point. You
> disagree with me on the change made here, but then I get a R-b? :) And
> then I go further down and read that the R-b has actually no value at
> all and I should wait for another one :D Maybe I'm overthinking this,
> but it didn't used to happen back when we used ACKs :)
>
>> with a few adjustments above; however, another R-By should be obtained
>> here as well as perhaps a policy change so that other similar such
>> series could be merged... I guess I'm curious what "thoughts"
others may
>> have regarding adding this "knob" while not allowing others.
>>
>> John
Did I disagree on the change or was I pointing out that from a certain
abstraction level it could be said the code is "similar to" the patches
I referenced w/r/t libvirt supplying some value that perhaps the
underlying hypervisor could/should have adjusted instead based on it's
knowledge of the rather involved algorithms. FWIW: I'm not opposed to
the referenced patches, but I'm in the minority on them.
AIUI, the R-b is essentially stating that I've looked at the patch and
in general agree with it. As part of that - there are a few things that
need adjustment and since I felt that it had some similarities to the
referenced patches w/r/t adding a knob that some may feel isn't
necessary I said let's make sure there's "enough consensus" that this
should be added (which it seems there is).
IIRC, the following was referenced at some point during the R-b discussion:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/D...
John