1 package PVE
::QemuServer
::Memory
;
6 use PVE
::Tools
qw(run_command lock_file lock_file_full file_read_firstline dir_glob_foreach);
7 use PVE
::Exception
qw(raise raise_param_exc);
12 my $MAX_MEM = 4194304;
15 sub get_numa_node_list
{
18 for (my $i = 0; $i < $MAX_NUMA; $i++) {
19 my $entry = $conf->{"numa$i"} or next;
20 my $numa = PVE
::QemuServer
::parse_numa
($entry) or next;
23 return @numa_map if @numa_map;
24 my $sockets = $conf->{sockets
} || 1;
25 return (0..($sockets-1));
28 # only valid when numa nodes map to a single host node
29 sub get_numa_guest_to_host_map
{
32 for (my $i = 0; $i < $MAX_NUMA; $i++) {
33 my $entry = $conf->{"numa$i"} or next;
34 my $numa = PVE
::QemuServer
::parse_numa
($entry) or next;
35 $map->{$i} = print_numa_hostnodes
($numa->{hostnodes
});
38 my $sockets = $conf->{sockets
} || 1;
39 return {map { $_ => $_ } (0..($sockets-1))};
43 my ($conf, $vmid, $memory, $sockets, $func) = @_;
49 if($conf->{hugepages
} && $conf->{hugepages
} == 1024) {
50 $current_size = 1024 * $sockets;
57 return if $current_size == $memory;
59 my @numa_map = get_numa_node_list
($conf);
61 for (my $j = 0; $j < 8; $j++) {
62 for (my $i = 0; $i < 32; $i++) {
63 my $name = "dimm${dimm_id}";
65 my $numanode = $numa_map[$i % @numa_map];
66 $current_size += $dimm_size;
67 &$func($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory);
68 return $current_size if $current_size >= $memory;
74 sub foreach_reverse_dimm
{
75 my ($conf, $vmid, $memory, $sockets, $func) = @_;
81 if($conf->{hugepages
} && $conf->{hugepages
} == 1024) {
82 $current_size = 8355840;
85 $current_size = 4177920;
89 return if $current_size == $memory;
91 my @numa_map = get_numa_node_list
($conf);
93 for (my $j = 0; $j < 8; $j++) {
94 for (my $i = 0; $i < 32; $i++) {
95 my $name = "dimm${dimm_id}";
97 my $numanode = $numa_map[(31-$i) % @numa_map];
98 $current_size -= $dimm_size;
99 &$func($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory);
100 return $current_size if $current_size <= $memory;
106 sub qemu_memory_hotplug
{
107 my ($vmid, $conf, $defaults, $opt, $value) = @_;
109 return $value if !PVE
::QemuServer
::check_running
($vmid);
112 $sockets = $conf->{sockets
} if $conf->{sockets
};
114 my $memory = $conf->{memory
} || $defaults->{memory
};
115 $value = $defaults->{memory
} if !$value;
116 return $value if $value == $memory;
118 my $static_memory = $STATICMEM;
119 $static_memory = $static_memory * $sockets if ($conf->{hugepages
} && $conf->{hugepages
} == 1024);
121 die "memory can't be lower than $static_memory MB" if $value < $static_memory;
122 die "you cannot add more memory than $MAX_MEM MB!\n" if $memory > $MAX_MEM;
124 if($value > $memory) {
126 my $numa_hostmap = get_numa_guest_to_host_map
($conf) if $conf->{hugepages
};
128 foreach_dimm
($conf, $vmid, $value, $sockets, sub {
129 my ($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory) = @_;
131 return if $current_size <= $conf->{memory
};
133 if ($conf->{hugepages
}) {
135 my $hugepages_size = hugepages_size
($conf, $dimm_size);
136 my $path = hugepages_mount_path
($hugepages_size);
137 my $host_numanode = $numa_hostmap->{$numanode};
138 my $hugepages_topology->{$hugepages_size}->{$host_numanode} = hugepages_nr
($dimm_size, $hugepages_size);
141 my $hugepages_host_topology = hugepages_host_topology
();
142 hugepages_allocate
($hugepages_topology, $hugepages_host_topology);
144 eval { PVE
::QemuServer
::vm_mon_cmd
($vmid, "object-add", 'qom-type' => "memory-backend-file", id
=> "mem-$name", props
=> {
145 size
=> int($dimm_size*1024*1024), 'mem-path' => $path, share
=> JSON
::true
, prealloc
=> JSON
::true
} ); };
147 hugepages_reset
($hugepages_host_topology);
151 hugepages_pre_deallocate
($hugepages_topology);
153 eval { hugepages_update_locked
($code); };
156 eval { PVE
::QemuServer
::vm_mon_cmd
($vmid, "object-add", 'qom-type' => "memory-backend-ram", id
=> "mem-$name", props
=> { size
=> int($dimm_size*1024*1024) } ) };
160 eval { PVE
::QemuServer
::qemu_objectdel
($vmid, "mem-$name"); };
164 eval { PVE
::QemuServer
::vm_mon_cmd
($vmid, "device_add", driver
=> "pc-dimm", id
=> "$name", memdev
=> "mem-$name", node
=> $numanode) };
166 eval { PVE
::QemuServer
::qemu_objectdel
($vmid, "mem-$name"); };
169 #update conf after each succesful module hotplug
170 $conf->{memory
} = $current_size;
171 PVE
::QemuConfig-
>write_config($vmid, $conf);
176 foreach_reverse_dimm
($conf, $vmid, $value, $sockets, sub {
177 my ($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory) = @_;
179 return if $current_size >= $conf->{memory
};
180 print "try to unplug memory dimm $name\n";
184 eval { PVE
::QemuServer
::qemu_devicedel
($vmid, $name) };
186 my $dimm_list = qemu_dimm_list
($vmid);
187 last if !$dimm_list->{$name};
188 raise_param_exc
({ $name => "error unplug memory module" }) if $retry > 5;
192 #update conf after each succesful module unplug
193 $conf->{memory
} = $current_size;
195 eval { PVE
::QemuServer
::qemu_objectdel
($vmid, "mem-$name"); };
196 PVE
::QemuConfig-
>write_config($vmid, $conf);
204 my $dimmarray = PVE
::QemuServer
::vm_mon_cmd_nocheck
($vmid, "query-memory-devices");
207 foreach my $dimm (@$dimmarray) {
209 $dimms->{$dimm->{data
}->{id
}}->{id
} = $dimm->{data
}->{id
};
210 $dimms->{$dimm->{data
}->{id
}}->{node
} = $dimm->{data
}->{node
};
211 $dimms->{$dimm->{data
}->{id
}}->{addr
} = $dimm->{data
}->{addr
};
212 $dimms->{$dimm->{data
}->{id
}}->{size
} = $dimm->{data
}->{size
};
213 $dimms->{$dimm->{data
}->{id
}}->{slot
} = $dimm->{data
}->{slot
};
219 my ($conf, $vmid, $sockets, $cores, $defaults, $hotplug_features, $cmd) = @_;
221 my $memory = $conf->{memory
} || $defaults->{memory
};
222 my $static_memory = 0;
224 if ($hotplug_features->{memory
}) {
225 die "NUMA needs to be enabled for memory hotplug\n" if !$conf->{numa
};
226 die "Total memory is bigger than ${MAX_MEM}MB\n" if $memory > $MAX_MEM;
228 $sockets = $conf->{sockets
} if $conf->{sockets
};
230 $static_memory = $STATICMEM;
231 $static_memory = $static_memory * $sockets if ($conf->{hugepages
} && $conf->{hugepages
} == 1024);
233 die "minimum memory must be ${static_memory}MB\n" if($memory < $static_memory);
234 push @$cmd, '-m', "size=${static_memory},slots=255,maxmem=${MAX_MEM}M";
238 $static_memory = $memory;
239 push @$cmd, '-m', $static_memory;
242 die "numa needs to be enabled to use hugepages" if $conf->{hugepages
} && !$conf->{numa
};
246 my $numa_totalmemory = undef;
247 for (my $i = 0; $i < $MAX_NUMA; $i++) {
248 next if !$conf->{"numa$i"};
249 my $numa = PVE
::QemuServer
::parse_numa
($conf->{"numa$i"});
252 die "missing NUMA node$i memory value\n" if !$numa->{memory
};
253 my $numa_memory = $numa->{memory
};
254 $numa_totalmemory += $numa_memory;
256 my $mem_object = print_mem_object
($conf, "ram-node$i", $numa_memory);
259 my $cpulists = $numa->{cpus
};
260 die "missing NUMA node$i cpus\n" if !defined($cpulists);
261 my $cpus = join(',cpus=', map {
262 my ($start, $end) = @$_;
263 defined($end) ?
"$start-$end" : $start
267 my $hostnodelists = $numa->{hostnodes
};
268 if (defined($hostnodelists)) {
270 my $hostnodes = print_numa_hostnodes
($hostnodelists);
273 my $policy = $numa->{policy
};
274 die "you need to define a policy for hostnode $hostnodes\n" if !$policy;
275 $mem_object .= ",host-nodes=$hostnodes,policy=$policy";
277 die "numa hostnodes need to be defined to use hugepages" if $conf->{hugepages
};
280 push @$cmd, '-object', $mem_object;
281 push @$cmd, '-numa', "node,nodeid=$i,cpus=$cpus,memdev=ram-node$i";
284 die "total memory for NUMA nodes must be equal to vm static memory\n"
285 if $numa_totalmemory && $numa_totalmemory != $static_memory;
287 #if no custom tology, we split memory and cores across numa nodes
288 if(!$numa_totalmemory) {
290 my $numa_memory = ($static_memory / $sockets);
292 for (my $i = 0; $i < $sockets; $i++) {
293 die "host NUMA node$i doesn't exist\n" if ! -d
"/sys/devices/system/node/node$i/" && $conf->{hugepages
};
295 my $cpustart = ($cores * $i);
296 my $cpuend = ($cpustart + $cores - 1) if $cores && $cores > 1;
297 my $cpus = $cpustart;
298 $cpus .= "-$cpuend" if $cpuend;
300 my $mem_object = print_mem_object
($conf, "ram-node$i", $numa_memory);
302 push @$cmd, '-object', $mem_object;
303 push @$cmd, '-numa', "node,nodeid=$i,cpus=$cpus,memdev=ram-node$i";
308 if ($hotplug_features->{memory
}) {
309 foreach_dimm
($conf, $vmid, $memory, $sockets, sub {
310 my ($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory) = @_;
312 my $mem_object = print_mem_object
($conf, "mem-$name", $dimm_size);
314 push @$cmd, "-object" , $mem_object;
315 push @$cmd, "-device", "pc-dimm,id=$name,memdev=mem-$name,node=$numanode";
317 #if dimm_memory is not aligned to dimm map
318 if($current_size > $memory) {
319 $conf->{memory
} = $current_size;
320 PVE
::QemuConfig-
>write_config($vmid, $conf);
326 sub print_mem_object
{
327 my ($conf, $id, $size) = @_;
329 if ($conf->{hugepages
}) {
331 my $hugepages_size = hugepages_size
($conf, $size);
332 my $path = hugepages_mount_path
($hugepages_size);
334 return "memory-backend-file,id=$id,size=${size}M,mem-path=$path,share=on,prealloc=yes";
336 return "memory-backend-ram,id=$id,size=${size}M";
341 sub print_numa_hostnodes
{
342 my ($hostnodelists) = @_;
345 foreach my $hostnoderange (@$hostnodelists) {
346 my ($start, $end) = @$hostnoderange;
347 $hostnodes .= ',' if $hostnodes;
348 $hostnodes .= $start;
349 $hostnodes .= "-$end" if defined($end);
351 for (my $i = $start; $i <= $end; ++$i ) {
352 die "host NUMA node$i doesn't exist\n" if ! -d
"/sys/devices/system/node/node$i/";
358 sub hugepages_mount
{
360 my $mountdata = PVE
::ProcFSTools
::parse_proc_mounts
();
362 foreach my $size (qw(2048 1048576)) {
363 next if (! -d
"/sys/kernel/mm/hugepages/hugepages-${size}kB");
365 my $path = "/run/hugepages/kvm/${size}kB";
368 $_->[2] =~ /^hugetlbfs/ &&
374 File
::Path
::make_path
($path) if (!-d
$path);
375 my $cmd = ['/bin/mount', '-t', 'hugetlbfs', '-o', "pagesize=${size}k", 'hugetlbfs', $path];
376 run_command
($cmd, errmsg
=> "hugepage mount error");
381 sub hugepages_mount_path
{
384 $size = $size * 1024;
385 return "/run/hugepages/kvm/${size}kB";
390 my ($size, $hugepages_size) = @_;
392 return $size / $hugepages_size;
396 my ($conf, $size) = @_;
397 die "hugepages option is not enabled" if !$conf->{hugepages
};
398 die "memory size '$size' is not a positive even integer; cannot use for hugepages\n"
399 if $size <= 0 || $size & 1;
401 my $page_chunk = sub { -d
"/sys/kernel/mm/hugepages/hugepages-". ($_[0] * 1024) ."kB" };
402 die "your system doesn't support hugepages\n" if !$page_chunk->(2) && !$page_chunk->(1024);
404 if ($conf->{hugepages
} eq 'any') {
406 # try to use 1GB if available && memory size is matching
407 if ($page_chunk->(1024) && ($size & 1023) == 0) {
409 } elsif ($page_chunk->(2)) {
412 die "host only supports 1024 GB hugepages, but requested size '$size' is not a multiple of 1024 MB\n"
416 my $hugepagesize = $conf->{hugepages
};
418 if (!$page_chunk->($hugepagesize)) {
419 die "your system doesn't support hugepages of $hugepagesize MB\n";
420 } elsif (($size % $hugepagesize) != 0) {
421 die "Memory size $size is not a multiple of the requested hugepages size $hugepagesize\n";
428 sub hugepages_topology
{
431 my $hugepages_topology = {};
433 return if !$conf->{numa
};
435 my $defaults = PVE
::QemuServer
::load_defaults
();
436 my $memory = $conf->{memory
} || $defaults->{memory
};
437 my $static_memory = 0;
439 $sockets = $conf->{smp
} if $conf->{smp
}; # old style - no longer iused
440 $sockets = $conf->{sockets
} if $conf->{sockets
};
441 my $numa_custom_topology = undef;
442 my $hotplug_features = PVE
::QemuServer
::parse_hotplug_features
(defined($conf->{hotplug
}) ?
$conf->{hotplug
} : '1');
444 if ($hotplug_features->{memory
}) {
445 $static_memory = $STATICMEM;
446 $static_memory = $static_memory * $sockets if ($conf->{hugepages
} && $conf->{hugepages
} == 1024);
448 $static_memory = $memory;
451 #custom numa topology
452 for (my $i = 0; $i < $MAX_NUMA; $i++) {
453 next if !$conf->{"numa$i"};
454 my $numa = PVE
::QemuServer
::parse_numa
($conf->{"numa$i"});
457 $numa_custom_topology = 1;
458 my $numa_memory = $numa->{memory
};
459 my $hostnodelists = $numa->{hostnodes
};
460 my $hostnodes = print_numa_hostnodes
($hostnodelists);
462 die "more than 1 hostnode value in numa node is not supported when hugepages are enabled" if $hostnodes !~ m/^(\d)$/;
463 my $hugepages_size = hugepages_size
($conf, $numa_memory);
464 $hugepages_topology->{$hugepages_size}->{$hostnodes} += hugepages_nr
($numa_memory, $hugepages_size);
468 #if no custom numa tology, we split memory and cores across numa nodes
469 if(!$numa_custom_topology) {
471 my $numa_memory = ($static_memory / $sockets);
473 for (my $i = 0; $i < $sockets; $i++) {
475 my $hugepages_size = hugepages_size
($conf, $numa_memory);
476 $hugepages_topology->{$hugepages_size}->{$i} += hugepages_nr
($numa_memory, $hugepages_size);
480 if ($hotplug_features->{memory
}) {
481 my $numa_hostmap = get_numa_guest_to_host_map
($conf);
483 foreach_dimm
($conf, undef, $memory, $sockets, sub {
484 my ($conf, undef, $name, $dimm_size, $numanode, $current_size, $memory) = @_;
486 $numanode = $numa_hostmap->{$numanode};
488 my $hugepages_size = hugepages_size
($conf, $dimm_size);
489 $hugepages_topology->{$hugepages_size}->{$numanode} += hugepages_nr
($dimm_size, $hugepages_size);
493 return $hugepages_topology;
496 sub hugepages_host_topology
{
499 my $hugepages_host_topology = {};
501 dir_glob_foreach
("/sys/devices/system/node/", 'node(\d+)', sub {
502 my ($nodepath, $numanode) = @_;
504 dir_glob_foreach
("/sys/devices/system/node/$nodepath/hugepages/", 'hugepages\-(\d+)kB', sub {
505 my ($hugepages_path, $hugepages_size) = @_;
507 $hugepages_size = $hugepages_size / 1024;
508 my $hugepages_nr = PVE
::Tools
::file_read_firstline
("/sys/devices/system/node/$nodepath/hugepages/$hugepages_path/nr_hugepages");
509 $hugepages_host_topology->{$hugepages_size}->{$numanode} = $hugepages_nr;
513 return $hugepages_host_topology;
516 sub hugepages_allocate
{
517 my ($hugepages_topology, $hugepages_host_topology) = @_;
519 #allocate new hupages if needed
520 foreach my $size (sort keys %$hugepages_topology) {
522 my $nodes = $hugepages_topology->{$size};
524 foreach my $numanode (keys %$nodes) {
526 my $hugepages_size = $size * 1024;
527 my $hugepages_requested = $hugepages_topology->{$size}->{$numanode};
528 my $path = "/sys/devices/system/node/node${numanode}/hugepages/hugepages-${hugepages_size}kB/";
529 my $hugepages_free = PVE
::Tools
::file_read_firstline
($path."free_hugepages");
530 my $hugepages_nr = PVE
::Tools
::file_read_firstline
($path."nr_hugepages");
532 if ($hugepages_requested > $hugepages_free) {
533 my $hugepages_needed = $hugepages_requested - $hugepages_free;
534 PVE
::ProcFSTools
::write_proc_entry
($path."nr_hugepages", $hugepages_nr + $hugepages_needed);
535 #verify that is correctly allocated
536 $hugepages_free = PVE
::Tools
::file_read_firstline
($path."free_hugepages");
537 if ($hugepages_free < $hugepages_requested) {
538 #rollback to initial host config
539 hugepages_reset
($hugepages_host_topology);
540 die "hugepage allocation failed";
549 sub hugepages_default_nr_hugepages
{
552 my $cmdline = PVE
::Tools
::file_read_firstline
("/proc/cmdline");
553 my $args = PVE
::Tools
::split_args
($cmdline);
555 my $parsed_size = 2; # default is 2M
557 foreach my $arg (@$args) {
558 if ($arg eq "hugepagesz=2M") {
560 } elsif ($arg eq "hugepagesz=1G") {
562 } elsif ($arg =~ m/^hugepages=(\d+)?$/) {
563 if ($parsed_size == $size) {
572 sub hugepages_pre_deallocate
{
573 my ($hugepages_topology) = @_;
575 foreach my $size (sort keys %$hugepages_topology) {
577 my $hugepages_size = $size * 1024;
578 my $path = "/sys/kernel/mm/hugepages/hugepages-${hugepages_size}kB/";
579 my $hugepages_nr = hugepages_default_nr_hugepages
($size);
580 PVE
::ProcFSTools
::write_proc_entry
($path."nr_hugepages", $hugepages_nr);
584 sub hugepages_reset
{
585 my ($hugepages_topology) = @_;
587 foreach my $size (sort keys %$hugepages_topology) {
589 my $nodes = $hugepages_topology->{$size};
590 foreach my $numanode (keys %$nodes) {
592 my $hugepages_nr = $hugepages_topology->{$size}->{$numanode};
593 my $hugepages_size = $size * 1024;
594 my $path = "/sys/devices/system/node/node${numanode}/hugepages/hugepages-${hugepages_size}kB/";
596 PVE
::ProcFSTools
::write_proc_entry
($path."nr_hugepages", $hugepages_nr);
601 sub hugepages_update_locked
{
602 my ($code, @param) = @_;
604 my $timeout = 60; #could be long if a lot of hugepages need to be alocated
606 my $lock_filename = "/var/lock/hugepages.lck";
608 my $res = lock_file
($lock_filename, $timeout, $code, @param);