Re: [pve-devel] [PATCH qemu-server 08/10] memory: add virtio-mem support

[Fiona Ebner <f.ebner@proxmox.com>]

Am 09.12.22 um 20:27 schrieb Alexandre Derumier:
> a 4GB static memory is needed for DMA+boot memory, as this memory
> is almost always un-unpluggeable.
> 
> 1 virtio-mem pci device is setup for each numa node on pci.4 bridge
> 
> virtio-mem use a fixed blocksize with 32000 blocks
> Blocksize is computed from the maxmemory-4096/32000 with a minimum of
> 2MB to map THP.
> (lower blocksize = more chance to unplug memory).
> 
> fixes:
> https://bugzilla.proxmox.com/show_bug.cgi?id=931

Comment 7 talks about Windows, and virtio-mem is not supported there at
the moment, so I don't think we should consider it fixed ;)

> https://bugzilla.proxmox.com/show_bug.cgi?id=2949
> Signed-off-by: Alexandre Derumier <aderumier@odiso.com>
> ---
>  PVE/QemuServer.pm        |  8 +++-
>  PVE/QemuServer/Memory.pm | 98 +++++++++++++++++++++++++++++++++++++---
>  PVE/QemuServer/PCI.pm    |  8 ++++
>  3 files changed, 106 insertions(+), 8 deletions(-)
> 
> diff --git a/PVE/QemuServer.pm b/PVE/QemuServer.pm
> index 0d5b550..43fab29 100644
> --- a/PVE/QemuServer.pm
> +++ b/PVE/QemuServer.pm
> @@ -285,6 +285,12 @@ my $memory_fmt = {
>          optional => 1,
>  	enum => [@max_memory_list],
>      },
> +    virtio => {
> +	description => "enable virtio-mem memory",

We really should mention that it's a technology preview, and that it
only works for Linux >=5.8 guests currently:
https://virtio-mem.gitlab.io/user-guide/#important-current-limitations

> +	type => 'boolean',
> +	optional => 1,
> +	default => 0,
> +    },
>  };
>  
>  my $meta_info_fmt = {
> @@ -3898,7 +3904,7 @@ sub config_to_command {
>  	push @$cmd, get_cpu_options($conf, $arch, $kvm, $kvm_off, $machine_version, $winversion, $gpu_passthrough);
>      }
>  
> -    PVE::QemuServer::Memory::config($conf, $vmid, $sockets, $cores, $defaults, $hotplug_features, $cmd);
> +    PVE::QemuServer::Memory::config($conf, $vmid, $sockets, $cores, $defaults, $hotplug_features, $cmd, $devices, $bridges, $arch, $machine_type);
>  
>      push @$cmd, '-S' if $conf->{freeze};
>  
> diff --git a/PVE/QemuServer/Memory.pm b/PVE/QemuServer/Memory.pm
> index 8bbbf07..70ab65a 100644
> --- a/PVE/QemuServer/Memory.pm
> +++ b/PVE/QemuServer/Memory.pm
> @@ -8,6 +8,8 @@ use PVE::Exception qw(raise raise_param_exc);
>  
>  use PVE::QemuServer;
>  use PVE::QemuServer::Monitor qw(mon_cmd);
> +use PVE::QemuServer::PCI qw(print_pci_addr);
> +
>  use base qw(Exporter);
>  
>  our @EXPORT_OK = qw(
> @@ -27,7 +29,9 @@ my sub get_static_mem {
>      my $static_memory = 0;
>      my $memory = PVE::QemuServer::parse_memory($conf->{memory});
>  
> -    if($memory->{max}) {
> +    if ($memory->{virtio}) {
> +	$static_memory = 4096;
> +    } elsif ($memory->{max}) {
>  	my $dimm_size = $memory->{max} / 64;
>  	#static mem can't be lower than 4G and lower than 1 dimmsize by socket
>  	$static_memory = $dimm_size * $sockets;
> @@ -102,6 +106,24 @@ my sub get_max_mem {
>      return $cpu_max_mem;
>  }
>  
> +my sub get_virtiomem_block_size {
> +    my ($conf) = @_;
> +
> +    my $MAX_MEM = get_max_mem($conf);
> +    my $static_memory = get_static_mem($conf);
> +    my $memory = get_current_memory($conf);
> +    #virtiomem can map 32000 block size. try to use lowerst blocksize, lower = more chance to unplug memory.

Style nit: line too long

> +    my $blocksize = ($MAX_MEM - $static_memory) / 32000;
> +    #round next power of 2
> +    $blocksize = 2**(int(log($blocksize)/log(2))+1);

What if log($blocksize)/log(2) is exactly an integer? Do we still want
to add 1 then? If not, please use ceil(...) instead of int(...)+1. Well,
I guess it can't happen in practice by what values are possible for
$MAX_MEM and $static_memory, but still.

> +    #2MB is the minimum to be aligned with THP
> +    $blocksize = 2 if $blocksize < 2;

Nit: $blocksize is at least 2**1 after the previous caluclation, so this
isn't really needed.

> +
> +    die "memory size need to be multiple of $blocksize MB when virtio-mem is enabled" if ($memory % $blocksize != 0);

Missing newline in error message.

Style nit: line too long

> +
> +    return $blocksize;
> +}
> +
>  sub get_current_memory{
>      my ($conf) = @_;
>  
> @@ -224,7 +246,41 @@ sub qemu_memory_hotplug {
>      my $MAX_MEM = get_max_mem($conf);
>      die "you cannot add more memory than max mem $MAX_MEM MB!\n" if $value > $MAX_MEM;
>  
> -    if ($value > $memory) {
> +    my $confmem = PVE::QemuServer::parse_memory($conf->{memory});
> +
> +    if ($confmem->{virtio}) {
> +	my $blocksize = get_virtiomem_block_size($conf);
> +	my $requested_size = ($value - $static_memory) / $sockets * 1024 * 1024;
> +	my $totalsize = $static_memory;
> +	my $err = undef;
> +
> +	for (my $i = 0; $i < $sockets; $i++)  {
> +
> +	    my $id = "virtiomem$i";
> +	    my $retry = 0;
> +	    mon_cmd($vmid, 'qom-set', path => "/machine/peripheral/$id", property => "requested-size", value => int($requested_size));

I'd eval the mon_cmd's and also catch errors there.

> +
> +	    my $size = 0;
> +	    while (1) {
> +		sleep 1;

If there really is no good alternative to this querying loop, I'd rather
issue the qom-set command for all virtio-mem devices first, and then do
the loop. Maybe also move the sleep to the end of the loop.

> +		$size = mon_cmd($vmid, 'qom-get', path => "/machine/peripheral/$id", property => "size");
> +		$err = 1 if $retry > 5;
> +		last if $size eq $requested_size || $retry > 5;

I think, $requested_size doesn't have to be a multiple of $sockets, so
this should rather be int($requested_size), which is what you set above.

> +		$retry++;
> +	    }
> +	    $totalsize += ($size / 1024 / 1024 );
> +	}
> +	#update conf after each succesfull change
> +	if($err) {

But this is only done in the error case, not after each successful change.

> +	    my $mem = { max => $MAX_MEM, virtio => 1};
> +	    $mem->{current} = $totalsize;

Nit: int($totalsize) just to be sure?

> +	    $conf->{memory} = PVE::QemuServer::print_memory($mem);
> +	    PVE::QemuConfig->write_config($vmid, $conf);
> +	    raise_param_exc({ 'memory' => "error modify virtio memory" }) if $err;

It's not necessarily a parameter issue, please use die instead.

> +	}
> +	return $totalsize;

The other branches don't (explicitly) return anything.

> +
> +    } elsif ($value > $memory) {
>  
>  	my $numa_hostmap;
>  
> @@ -324,14 +380,15 @@ sub qemu_dimm_list {
>  }
>  
>  sub config {
> -    my ($conf, $vmid, $sockets, $cores, $defaults, $hotplug_features, $cmd) = @_;
> +    my ($conf, $vmid, $sockets, $cores, $defaults, $hotplug_features, $cmd, $devices, $bridges, $arch, $machine_type) = @_;
>  
>      my $memory = get_current_memory($conf);
>  
>      my $static_memory = get_static_mem($conf);
> +
>      my $confmem = PVE::QemuServer::parse_memory($conf->{memory});
>  
> -    if ($hotplug_features->{memory} || defined($confmem->{max})) {
> +    if ($hotplug_features->{memory} || defined($confmem->{max}) || defined($confmem->{virtio})) {

Again, should we even bother attaching the devices if hotplug is not
enabled?

>  	die "NUMA needs to be enabled for memory hotplug\n" if !$conf->{numa};
>  	my $MAX_MEM = get_max_mem($conf);
>  	die "Total memory is bigger than ${MAX_MEM}MB\n" if $memory > $MAX_MEM;
> @@ -342,8 +399,12 @@ sub config {
>  	}
>  
>  	die "minimum memory must be ${static_memory}MB\n" if($memory < $static_memory);
> +
> +	my $cmdstr = "size=${static_memory}";
>  	my $slots = $confmem->{max} ? 64 : 255;
> -	push @$cmd, '-m', "size=${static_memory},slots=$slots,maxmem=${MAX_MEM}M";
> +	$cmdstr .= ",slots=$slots" if !$confmem->{'virtio'};
> +	$cmdstr .= ",maxmem=${MAX_MEM}M";
> +	push @$cmd, '-m', $cmdstr;
>  
>      } else {
>  	push @$cmd, '-m', $static_memory;
> @@ -412,7 +473,26 @@ sub config {
>  	}
>      }
>  
> -    if ($hotplug_features->{memory} || $confmem->{max}) {
> +    if ($confmem->{'virtio'}) {
> +	my $MAX_MEM = get_max_mem($conf);
> +	my $node_maxmem = ($MAX_MEM - $static_memory) / $sockets;
> +	my $node_mem = ($memory - $static_memory) / $sockets;
> +	my $blocksize = get_virtiomem_block_size($conf);
> +
> +	for (my $i = 0; $i < $sockets; $i++)  {
> +
> +	    my $id = "virtiomem$i";
> +	    my $mem_object = print_mem_object($conf, "mem-$id", $node_maxmem);
> +	    push @$cmd, "-object" , "$mem_object,reserve=off";
> +
> +	    my $pciaddr = print_pci_addr($id, $bridges, $arch, $machine_type);

Can we rather handle the PCI address printing in config_to_command() and
only pass in the addresses here? That would also avoid the PCI module
usage and the new "one-time usage" parameters passed to Memory::config().

Maybe have a small helper in here, that just returns the needed $ids
depending on the config. Then in config_to_command(), call that helper,
print the PCI addresses, then call Memory::config(..., { $id1 =>
$address1, $id2 => $adress2, ... }).

Might also not be the nicest, but at least be a little less cluttering
IMHO. But feel free to come up with something better or keep it as-is if
you really want to ;)

> +	    my $mem_device = "virtio-mem-pci,block-size=${blocksize}M,requested-size=${node_mem}M,id=$id,memdev=mem-$id,node=$i$pciaddr";
> +	    $mem_device .= ",prealloc=on" if $conf->{hugepages};

So prealloc=on for the device, but not prealloc=yes for the object
below[0]. Would you mind explaining it to me? I just found the part
mentioned for v7.0 here
https://virtio-mem.gitlab.io/user-guide/user-guide-qemu.html#updates

> +	    push @$devices, "-device", $mem_device;

The dimm devices in the other branch are not pushed onto $devices, so
this feels inconsistent. Why not add it onto $cmd too?

> +	}
> +
> +    } elsif ($hotplug_features->{memory} || $confmem->{max}) {
> +
>  	foreach_dimm($conf, $vmid, $memory, $sockets, sub {
>  	    my ($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory) = @_;
>  
> @@ -430,12 +510,16 @@ sub config {
>  sub print_mem_object {
>      my ($conf, $id, $size) = @_;
>  
> +    my $confmem = PVE::QemuServer::parse_memory($conf->{memory});
> +
>      if ($conf->{hugepages}) {
>  
>  	my $hugepages_size = hugepages_size($conf, $size);
>  	my $path = hugepages_mount_path($hugepages_size);
>  
> -	return "memory-backend-file,id=$id,size=${size}M,mem-path=$path,share=on,prealloc=yes";
> +	my $object = "memory-backend-file,id=$id,size=${size}M,mem-path=$path,share=on";
> +	$object .= ",prealloc=yes" if !$confmem->{virtio};

[0]

> +	return $object;
>      } else {
>  	return "memory-backend-ram,id=$id,size=${size}M";
>      }
> diff --git a/PVE/QemuServer/PCI.pm b/PVE/QemuServer/PCI.pm
> index a18b974..0187c74 100644
> --- a/PVE/QemuServer/PCI.pm
> +++ b/PVE/QemuServer/PCI.pm
> @@ -249,6 +249,14 @@ sub get_pci_addr_map {
>  	'scsihw2' => { bus => 4, addr => 1 },
>  	'scsihw3' => { bus => 4, addr => 2 },
>  	'scsihw4' => { bus => 4, addr => 3 },
> +	'virtiomem0' => { bus => 4, addr => 4 },
> +	'virtiomem1' => { bus => 4, addr => 5 },
> +	'virtiomem2' => { bus => 4, addr => 6 },
> +	'virtiomem3' => { bus => 4, addr => 7 },
> +	'virtiomem4' => { bus => 4, addr => 8 },
> +	'virtiomem5' => { bus => 4, addr => 9 },
> +	'virtiomem6' => { bus => 4, addr => 10 },
> +	'virtiomem7' => { bus => 4, addr => 11 },

What if $conf->{sockets} > 8? Maybe mention the limitation in the
description of the 'virtio' property in the 'memory' string. Is the plan
to just add on more virtiomem PCI devices in the future?

>      } if !defined($pci_addr_map);
>      return $pci_addr_map;
>  }