[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Xen-devel] [PATCHv1] xen/balloon: disable memory hotplug in PV guests
On 03/17/2015 01:40 PM, Daniel Kiper wrote:
On Mon, Mar 16, 2015 at 11:31:49AM +0100, Juergen Gross wrote:
On 03/16/2015 11:03 AM, Daniel Kiper wrote:
On Mon, Mar 16, 2015 at 06:35:04AM +0100, Juergen Gross wrote:
On 03/11/2015 04:40 PM, Boris Ostrovsky wrote:
On 03/11/2015 10:42 AM, David Vrabel wrote:
On 10/03/15 13:35, Boris Ostrovsky wrote:
On 03/10/2015 07:40 AM, David Vrabel wrote:
On 09/03/15 14:10, David Vrabel wrote:
Memory hotplug doesn't work with PV guests because:
a) The p2m cannot be expanded to cover the new sections.
Broken by 054954eb051f35e74b75a566a96fe756015352c8 (xen: switch to
linear virtual mapped sparse p2m list).
This one would be non-trivial to fix. We'd need a sparse set of
vm_area's for the p2m or similar.
b) add_memory() builds page tables for the new sections which
means
the new pages must have valid p2m entries (or a BUG occurs).
After some more testing this appears to be broken by:
25b884a83d487fd62c3de7ac1ab5549979188482 (x86/xen: set regions above
the
end of RAM as 1:1) included 3.16.
This one can be trivially fixed by setting the new sections in the p2m
to INVALID_P2M_ENTRY before calling add_memory().
Have you tried 3.17? As I said yesterday, it worked for me (with 4.4
Xen).
No. But there are three bugs that prevent it from working in 3.16+ so
I'm really not sure how you had a working in a 3.17 PV guest.
This is what I have:
[build@build-mk2 linux-boris]$ ssh root@tst008 cat
/mnt/lab/bootstrap-x86_64/test_small.xm
extra="console=hvc0 debug earlyprintk=xen "
kernel="/mnt/lab/bootstrap-x86_64/vmlinuz"
ramdisk="/mnt/lab/bootstrap-x86_64/initramfs.cpio.gz"
memory=1024
maxmem = 4096
vcpus=1
maxvcpus=3
name="bootstrap-x86_64"
on_crash="preserve"
vif = [ 'mac=00:0F:4B:00:00:68, bridge=switch' ]
vnc=1
vnclisten="0.0.0.0"
disk=['phy:/dev/guests/bootstrap-x86_64,xvda,w']
[build@build-mk2 linux-boris]$ ssh root@tst008 xl create
/mnt/lab/bootstrap-x86_64/test_small.xm
Parsing config from /mnt/lab/bootstrap-x86_64/test_small.xm
[build@build-mk2 linux-boris]$ ssh root@tst008 xl list |grep
bootstrap-x86_64
bootstrap-x86_64 2 1024 1 -b---- 5.4
[build@build-mk2 linux-boris]$ ssh root@g-pvops uname -r
3.17.0upstream
[build@build-mk2 linux-boris]$ ssh root@g-pvops dmesg|grep paravirtualized
[ 0.000000] Booting paravirtualized kernel on Xen
[build@build-mk2 linux-boris]$ ssh root@g-pvops grep MemTotal /proc/meminfo
MemTotal: 968036 kB
[build@build-mk2 linux-boris]$ ssh root@tst008 xl mem-set
bootstrap-x86_64 2048
[build@build-mk2 linux-boris]$ ssh root@tst008 xl list |grep
bootstrap-x86_64
bootstrap-x86_64 2 2048 1 -b---- 5.7
[build@build-mk2 linux-boris]$ ssh root@g-pvops grep MemTotal /proc/meminfo
MemTotal: 2016612 kB
[build@build-mk2 linux-boris]$
Regardless, it definitely doesn't work now because of the linear p2m
change. What do you want to do about this?
Since backing out p2m changes is not an option I guess your patch is the
only short-term alternative.
But this still looks like a regression so perhaps Juergen can take a
look to see how it can be fixed.
Hmm, the p2m list is allocated for the maximum memory size of the domain
which is obtained from the hypervisor. In case of Dom0 it is read via
XENMEM_maximum_reservation, for a domU it is based on the E820 memory
map read via XENMEM_memory_map.
I just tested it with a 4.0-rc1 domU kernel with 512MB initial memory
and 4GB of maxmem. The E820 map looked like this:
[ 0.000000] Xen: [mem 0x0000000000000000-0x000000000009ffff] usable
[ 0.000000] Xen: [mem 0x00000000000a0000-0x00000000000fffff] reserved
[ 0.000000] Xen: [mem 0x0000000000100000-0x00000000ffffffff] usable
So the complete 4GB were included, like they should. The resulting p2m
list is allocated in the needed size:
[ 0.000000] p2m virtual area at ffffc90000000000, size is 800000
So what is your problem here? Can you post the E820 map and the p2m map
info for your failing domain, please?
If you use memory hotplug then maxmem is not a limit from guest kernel
point of view (host still must allow that operation but it is another
not related issue). The problem is that p2m must be dynamically expendable
to support it. Earlier implementation supported that thing and memory
hotplug worked without any issue.
Okay, now I get it.
The problem with the earlier p2m implementation was that it was
expendable to support only up to 512GB of RAM. So we need some way to
tell the kernel how much virtual memory it should reserve for the p2m
list if memory hotplug is enabled. We could:
a) use a configurable maximum (e.g. for 512GB RAM as today)
b) use the maximum of RAM the machine the domain is started on can ever
have (what about migration then?)
Memory hotplug for Xen is planned, so, this would not work.
c) use a kernel parameter specifying the maximum memory size to support
d) a combination of some of the above possibilities
Any thoughts? I think I'd prefer b)+c).
I do not know new p2m design well so correct me if I am wrong.
If you set size limit on p2m then memory hotplug does not make sense.
Ballooning solves all your problems in simple manner. Memory hotplug
is a solution if you would like to set up a guest with small amount
of memory, you are not able to predict future requirements (usually
you are not) and you are not able to restart machine at any given
moment due to tight requirements. So, I think that p2m size, if must
be limited, then it should be limited by platform limitations (e.g.
number of address lines) but not artificially as you suggested.
I was talking about platform limitations, but more like "maximum of RAM
which can be plugged into the current configuration". It makes no sense
to use a limit of TBs when the system has only 1 socket with 8 possible
DIMMs.
Additionally, I have a feeling that you are going to preallocate space
in p2m for potential memory hotplug uses. Am I right? If yes then
I think this is unneeded waste of memory.
I'm preallocating _virtual_ memory, not physical. There is some overhead
regarding physical memory, but this is very very small (about 4kB per
1GB of supported memory size). In case this small amount is really
critical, I could insert another optimization reducing the overhead by
another factor of 512.
As I know new p2m is a table with MFNs. If yes, then I think we can
potentially add more space for new MFNs at the and of that table (or
As it is virtual memory, this is already the case.
create new larger one then copy old table contents and destroy former
one; Hmmm... What about memory fragmentation which may prevent us
allocating this big chunk of memory...). If this works then this
process could be repeated without limitations (Well... We will be
just limited by platform itself).
Rebuilding the p2m list would work, but only if enough contiguous
virtual memory is available.
One problem remains: 32 bit domains are short of virtual memory space.
Do we need memory hotplug for 32 bit domains as well?
Juergen
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel
|
