Linux Virtualization - Jun 2018 - [PATCH v34 2/4] virtio-balloon: VIRTIO_BALLOON_F_FREE_PAGE_HINT

On Wed, Jun 27, 2018 at 11:00:05AM +0800, Wei Wang
wrote:
> On 06/27/2018 10:41 AM, Michael S. Tsirkin wrote:
> > On Wed, Jun 27, 2018 at 09:24:18AM +0800, Wei Wang wrote:
> > > On 06/26/2018 09:34 PM, Michael S. Tsirkin wrote:
> > > > On Tue, Jun 26, 2018 at 08:27:44PM +0800, Wei Wang wrote:
> > > > > On 06/26/2018 11:56 AM, Michael S. Tsirkin wrote:
> > > > > > On Tue, Jun 26, 2018 at 11:46:35AM +0800, Wei Wang
wrote:
> > > > > > 
> > > > > > > > > +	if (!arrays)
> > > > > > > > > +		return NULL;
> > > > > > > > > +
> > > > > > > > > +	for (i = 0; i < max_array_num;
i++) {
> > > > > > > > So we are getting a ton of memory here
just to free it up a bit later.
> > > > > > > > Why doesn't get_from_free_page_list
get the pages from free list for us?
> > > > > > > > We could also avoid the 1st allocation
then - just build a list
> > > > > > > > of these.
> > > > > > > That wouldn't be a good choice for us. If
we check how the regular
> > > > > > > allocation works, there are many many things
we need to consider when pages
> > > > > > > are allocated to users.
> > > > > > > For example, we need to take care of the
nr_free
> > > > > > > counter, we need to check the watermark and
perform the related actions.
> > > > > > > Also the folks working on arch_alloc_page to
monitor page allocation
> > > > > > > activities would get a surprise..if page
allocation is allowed to work in
> > > > > > > this way.
> > > > > > > 
> > > > > > mm/ code is well positioned to handle all this
correctly.
> > > > > I'm afraid that would be a re-implementation of the
alloc functions,
> > > > A re-factoring - you can share code. The main difference is
locking.
> > > > 
> > > > > and
> > > > > that would be much more complex than what we have. I
think your idea of
> > > > > passing a list of pages is better.
> > > > > 
> > > > > Best,
> > > > > Wei
> > > > How much memory is this allocating anyway?
> > > > 
> > > For every 2TB memory that the guest has, we allocate 4MB.
> > Hmm I guess I'm missing something, I don't see it:
> > 
> > 
> > +       max_entries = max_free_page_blocks(ARRAY_ALLOC_ORDER);
> > +       entries_per_page = PAGE_SIZE / sizeof(__le64);
> > +       entries_per_array = entries_per_page * (1 <<
ARRAY_ALLOC_ORDER);
> > +       max_array_num = max_entries / entries_per_array +
> > +                       !!(max_entries % entries_per_array);
> > 
> > Looks like you always allocate the max number?
> 
> Yes. We allocated the max number and then free what's not used.
> For example, a 16TB guest, we allocate Four 4MB buffers and pass the 4
> buffers to get_from_free_page_list. If it uses 3, then the remaining 1
"4MB
> buffer" will end up being freed.
> 
> For today's guests, max_array_num is usually 1.
> 
> Best,
> Wei
I see, it's based on total ram pages. It's reasonable but might
get out of sync if memory is onlined quickly. So you want to
detect that there's more free memory than can fit and
retry the reporting.
> 
> 
>

On 06/27/2018 11:58 AM, Michael S. Tsirkin wrote:
> On Wed, Jun 27, 2018 at 11:00:05AM +0800, Wei Wang wrote:
>> On 06/27/2018 10:41 AM, Michael S. Tsirkin wrote:
>>> On Wed, Jun 27, 2018 at 09:24:18AM +0800, Wei Wang wrote:
>>>> On 06/26/2018 09:34 PM, Michael S. Tsirkin wrote:
>>>>> On Tue, Jun 26, 2018 at 08:27:44PM +0800, Wei Wang wrote:
>>>>>> On 06/26/2018 11:56 AM, Michael S. Tsirkin wrote:
>>>>>>> On Tue, Jun 26, 2018 at 11:46:35AM +0800, Wei Wang
wrote:
>>>>>>>
>>>>>>>>>> +	if (!arrays)
>>>>>>>>>> +		return NULL;
>>>>>>>>>> +
>>>>>>>>>> +	for (i = 0; i < max_array_num;
i++) {
>>>>>>>>> So we are getting a ton of memory here just
to free it up a bit later.
>>>>>>>>> Why doesn't get_from_free_page_list get
the pages from free list for us?
>>>>>>>>> We could also avoid the 1st allocation then
- just build a list
>>>>>>>>> of these.
>>>>>>>> That wouldn't be a good choice for us. If
we check how the regular
>>>>>>>> allocation works, there are many many things we
need to consider when pages
>>>>>>>> are allocated to users.
>>>>>>>> For example, we need to take care of the
nr_free
>>>>>>>> counter, we need to check the watermark and
perform the related actions.
>>>>>>>> Also the folks working on arch_alloc_page to
monitor page allocation
>>>>>>>> activities would get a surprise..if page
allocation is allowed to work in
>>>>>>>> this way.
>>>>>>>>
>>>>>>> mm/ code is well positioned to handle all this
correctly.
>>>>>> I'm afraid that would be a re-implementation of the
alloc functions,
>>>>> A re-factoring - you can share code. The main difference is
locking.
>>>>>
>>>>>> and
>>>>>> that would be much more complex than what we have. I
think your idea of
>>>>>> passing a list of pages is better.
>>>>>>
>>>>>> Best,
>>>>>> Wei
>>>>> How much memory is this allocating anyway?
>>>>>
>>>> For every 2TB memory that the guest has, we allocate 4MB.
>>> Hmm I guess I'm missing something, I don't see it:
>>>
>>>
>>> +       max_entries = max_free_page_blocks(ARRAY_ALLOC_ORDER);
>>> +       entries_per_page = PAGE_SIZE / sizeof(__le64);
>>> +       entries_per_array = entries_per_page * (1 <<
ARRAY_ALLOC_ORDER);
>>> +       max_array_num = max_entries / entries_per_array +
>>> +                       !!(max_entries % entries_per_array);
>>>
>>> Looks like you always allocate the max number?
>> Yes. We allocated the max number and then free what's not used.
>> For example, a 16TB guest, we allocate Four 4MB buffers and pass the 4
>> buffers to get_from_free_page_list. If it uses 3, then the remaining 1
"4MB
>> buffer" will end up being freed.
>>
>> For today's guests, max_array_num is usually 1.
>>
>> Best,
>> Wei
> I see, it's based on total ram pages. It's reasonable but might
> get out of sync if memory is onlined quickly. So you want to
> detect that there's more free memory than can fit and
> retry the reporting.
>

- AFAIK, memory hotplug isn't expected to happen during live migration 
today. Hypervisors (e.g. QEMU) explicitly forbid this.

- Allocating buffers based on total ram pages already gives some 
headroom for newly plugged memory if that could happen in any case. 
Also, we can think about why people plug in more memory - usually 
because the existing memory isn't enough, which implies that the free 
page list is very likely to be close to empty.

- This method could be easily scaled if people really need more headroom 
for hot-plugged memory. For example, calculation based on "X * 
total_ram_pages", X could be a number passed from the hypervisor.

- This is an optimization feature, and reporting less free memory in 
that rare case doesn't hurt anything.

So I think it is good to start from a fundamental implementation, which 
doesn't confuse people, and complexities can be added when there is a 
real need in the future.

Best,
Wei

On Wed, Jun 27, 2018 at 01:27:55PM +0800, Wei Wang
wrote:
> On 06/27/2018 11:58 AM, Michael S. Tsirkin wrote:
> > On Wed, Jun 27, 2018 at 11:00:05AM +0800, Wei Wang wrote:
> > > On 06/27/2018 10:41 AM, Michael S. Tsirkin wrote:
> > > > On Wed, Jun 27, 2018 at 09:24:18AM +0800, Wei Wang wrote:
> > > > > On 06/26/2018 09:34 PM, Michael S. Tsirkin wrote:
> > > > > > On Tue, Jun 26, 2018 at 08:27:44PM +0800, Wei Wang
wrote:
> > > > > > > On 06/26/2018 11:56 AM, Michael S. Tsirkin
wrote:
> > > > > > > > On Tue, Jun 26, 2018 at 11:46:35AM
+0800, Wei Wang wrote:
> > > > > > > > 
> > > > > > > > > > > +	if (!arrays)
> > > > > > > > > > > +		return NULL;
> > > > > > > > > > > +
> > > > > > > > > > > +	for (i = 0; i <
max_array_num; i++) {
> > > > > > > > > > So we are getting a ton of
memory here just to free it up a bit later.
> > > > > > > > > > Why doesn't
get_from_free_page_list get the pages from free list for us?
> > > > > > > > > > We could also avoid the 1st
allocation then - just build a list
> > > > > > > > > > of these.
> > > > > > > > > That wouldn't be a good choice
for us. If we check how the regular
> > > > > > > > > allocation works, there are many
many things we need to consider when pages
> > > > > > > > > are allocated to users.
> > > > > > > > > For example, we need to take care
of the nr_free
> > > > > > > > > counter, we need to check the
watermark and perform the related actions.
> > > > > > > > > Also the folks working on
arch_alloc_page to monitor page allocation
> > > > > > > > > activities would get a surprise..if
page allocation is allowed to work in
> > > > > > > > > this way.
> > > > > > > > > 
> > > > > > > > mm/ code is well positioned to handle
all this correctly.
> > > > > > > I'm afraid that would be a
re-implementation of the alloc functions,
> > > > > > A re-factoring - you can share code. The main
difference is locking.
> > > > > > 
> > > > > > > and
> > > > > > > that would be much more complex than what we
have. I think your idea of
> > > > > > > passing a list of pages is better.
> > > > > > > 
> > > > > > > Best,
> > > > > > > Wei
> > > > > > How much memory is this allocating anyway?
> > > > > > 
> > > > > For every 2TB memory that the guest has, we allocate
4MB.
> > > > Hmm I guess I'm missing something, I don't see it:
> > > > 
> > > > 
> > > > +       max_entries =
max_free_page_blocks(ARRAY_ALLOC_ORDER);
> > > > +       entries_per_page = PAGE_SIZE / sizeof(__le64);
> > > > +       entries_per_array = entries_per_page * (1 <<
ARRAY_ALLOC_ORDER);
> > > > +       max_array_num = max_entries / entries_per_array +
> > > > +                       !!(max_entries % entries_per_array);
> > > > 
> > > > Looks like you always allocate the max number?
> > > Yes. We allocated the max number and then free what's not
used.
> > > For example, a 16TB guest, we allocate Four 4MB buffers and pass
the 4
> > > buffers to get_from_free_page_list. If it uses 3, then the
remaining 1 "4MB
> > > buffer" will end up being freed.
> > > 
> > > For today's guests, max_array_num is usually 1.
> > > 
> > > Best,
> > > Wei
> > I see, it's based on total ram pages. It's reasonable but
might
> > get out of sync if memory is onlined quickly. So you want to
> > detect that there's more free memory than can fit and
> > retry the reporting.
> > 
> 
> 
> - AFAIK, memory hotplug isn't expected to happen during live migration
> today. Hypervisors (e.g. QEMU) explicitly forbid this.
That's a temporary limitation.
> - Allocating buffers based on total ram pages already gives some headroom
> for newly plugged memory if that could happen in any case. Also, we can
> think about why people plug in more memory - usually because the existing
> memory isn't enough, which implies that the free page list is very
likely to
> be close to empty.
Or maybe because guest is expected to use more memory.
> - This method could be easily scaled if people really need more headroom
for
> hot-plugged memory. For example, calculation based on "X *
total_ram_pages",
> X could be a number passed from the hypervisor.
All this in place of a simple retry loop within guest?
> - This is an optimization feature, and reporting less free memory in that
> rare case doesn't hurt anything.
People working on memory hotplug can't be expected to worry about
balloon. And maintainers have other things to do than debug hard to
trigger failure reports from the field.
> 
> So I think it is good to start from a fundamental implementation, which
> doesn't confuse people, and complexities can be added when there is a
real
> need in the future.
> 
> Best,
> Wei
The usefulness of the whole patchset hasn't been proven in the field yet.
The more uncovered corner cases there are, the higher the chance that
it will turn out not to be useful after all.
> 
> 
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: virtio-dev-unsubscribe at lists.oasis-open.org
> For additional commands, e-mail: virtio-dev-help at lists.oasis-open.org