]>
Commit | Line | Data |
---|---|---|
a0649da2 AD |
1 | package PVE::QemuServer::Memory; |
2 | ||
3 | use strict; | |
4 | use warnings; | |
71aba4ea | 5 | |
7023f3ea | 6 | use PVE::Tools qw(run_command lock_file lock_file_full file_read_firstline dir_glob_foreach); |
9a1c413f | 7 | use PVE::Exception qw(raise raise_param_exc); |
a0649da2 | 8 | |
71aba4ea TL |
9 | use PVE::QemuServer; |
10 | ||
0567a4d5 | 11 | my $MAX_NUMA = 8; |
6779f1ac AD |
12 | my $MAX_MEM = 4194304; |
13 | my $STATICMEM = 1024; | |
a0649da2 | 14 | |
aaff69ad WB |
15 | sub get_numa_node_list { |
16 | my ($conf) = @_; | |
17 | my @numa_map; | |
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; | |
21 | push @numa_map, $i; | |
22 | } | |
23 | return @numa_map if @numa_map; | |
24 | my $sockets = $conf->{sockets} || 1; | |
25 | return (0..($sockets-1)); | |
26 | } | |
27 | ||
28 | # only valid when numa nodes map to a single host node | |
29 | sub get_numa_guest_to_host_map { | |
30 | my ($conf) = @_; | |
31 | my $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}); | |
36 | } | |
37 | return $map if %$map; | |
38 | my $sockets = $conf->{sockets} || 1; | |
1d682951 | 39 | return {map { $_ => $_ } (0..($sockets-1))}; |
aaff69ad WB |
40 | } |
41 | ||
3f669af2 AD |
42 | sub foreach_dimm{ |
43 | my ($conf, $vmid, $memory, $sockets, $func) = @_; | |
44 | ||
45 | my $dimm_id = 0; | |
7023f3ea AD |
46 | my $current_size = 0; |
47 | my $dimm_size = 0; | |
48 | ||
49 | if($conf->{hugepages} && $conf->{hugepages} == 1024) { | |
50 | $current_size = 1024 * $sockets; | |
51 | $dimm_size = 1024; | |
52 | } else { | |
53 | $current_size = 1024; | |
54 | $dimm_size = 512; | |
55 | } | |
56 | ||
3f669af2 AD |
57 | return if $current_size == $memory; |
58 | ||
aaff69ad WB |
59 | my @numa_map = get_numa_node_list($conf); |
60 | ||
3f669af2 AD |
61 | for (my $j = 0; $j < 8; $j++) { |
62 | for (my $i = 0; $i < 32; $i++) { | |
63 | my $name = "dimm${dimm_id}"; | |
64 | $dimm_id++; | |
aaff69ad | 65 | my $numanode = $numa_map[$i % @numa_map]; |
3f669af2 AD |
66 | $current_size += $dimm_size; |
67 | &$func($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory); | |
68 | return $current_size if $current_size >= $memory; | |
69 | } | |
70 | $dimm_size *= 2; | |
71 | } | |
72 | } | |
73 | ||
74 | sub foreach_reverse_dimm { | |
75 | my ($conf, $vmid, $memory, $sockets, $func) = @_; | |
76 | ||
77 | my $dimm_id = 253; | |
7023f3ea AD |
78 | my $current_size = 0; |
79 | my $dimm_size = 0; | |
80 | ||
81 | if($conf->{hugepages} && $conf->{hugepages} == 1024) { | |
82 | $current_size = 8355840; | |
83 | $dimm_size = 131072; | |
84 | } else { | |
85 | $current_size = 4177920; | |
86 | $dimm_size = 65536; | |
87 | } | |
88 | ||
3f669af2 AD |
89 | return if $current_size == $memory; |
90 | ||
aaff69ad WB |
91 | my @numa_map = get_numa_node_list($conf); |
92 | ||
3f669af2 AD |
93 | for (my $j = 0; $j < 8; $j++) { |
94 | for (my $i = 0; $i < 32; $i++) { | |
95 | my $name = "dimm${dimm_id}"; | |
96 | $dimm_id--; | |
aaff69ad | 97 | my $numanode = $numa_map[(31-$i) % @numa_map]; |
3f669af2 AD |
98 | $current_size -= $dimm_size; |
99 | &$func($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory); | |
100 | return $current_size if $current_size <= $memory; | |
101 | } | |
102 | $dimm_size /= 2; | |
103 | } | |
104 | } | |
105 | ||
6779f1ac AD |
106 | sub qemu_memory_hotplug { |
107 | my ($vmid, $conf, $defaults, $opt, $value) = @_; | |
108 | ||
37dbec89 | 109 | return $value if !PVE::QemuServer::check_running($vmid); |
6779f1ac | 110 | |
7023f3ea AD |
111 | my $sockets = 1; |
112 | $sockets = $conf->{sockets} if $conf->{sockets}; | |
113 | ||
6779f1ac AD |
114 | my $memory = $conf->{memory} || $defaults->{memory}; |
115 | $value = $defaults->{memory} if !$value; | |
116 | return $value if $value == $memory; | |
117 | ||
118 | my $static_memory = $STATICMEM; | |
7023f3ea | 119 | $static_memory = $static_memory * $sockets if ($conf->{hugepages} && $conf->{hugepages} == 1024); |
6779f1ac AD |
120 | |
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 > $MAX_MEM; | |
123 | ||
6779f1ac AD |
124 | if($value > $memory) { |
125 | ||
aaff69ad WB |
126 | my $numa_hostmap = get_numa_guest_to_host_map($conf) if $conf->{hugepages}; |
127 | ||
6779f1ac AD |
128 | foreach_dimm($conf, $vmid, $value, $sockets, sub { |
129 | my ($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory) = @_; | |
130 | ||
131 | return if $current_size <= $conf->{memory}; | |
132 | ||
7023f3ea AD |
133 | if ($conf->{hugepages}) { |
134 | ||
135 | my $hugepages_size = hugepages_size($conf, $dimm_size); | |
136 | my $path = hugepages_mount_path($hugepages_size); | |
aaff69ad WB |
137 | my $host_numanode = $numa_hostmap->{$numanode}; |
138 | my $hugepages_topology->{$hugepages_size}->{$host_numanode} = hugepages_nr($dimm_size, $hugepages_size); | |
7023f3ea AD |
139 | |
140 | my $code = sub { | |
141 | my $hugepages_host_topology = hugepages_host_topology(); | |
142 | hugepages_allocate($hugepages_topology, $hugepages_host_topology); | |
143 | ||
144 | eval { PVE::QemuServer::vm_mon_cmd($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 } ); }; | |
146 | if (my $err = $@) { | |
147 | hugepages_reset($hugepages_host_topology); | |
148 | die $err; | |
149 | } | |
150 | ||
151 | hugepages_pre_deallocate($hugepages_topology); | |
152 | }; | |
153 | eval { hugepages_update_locked($code); }; | |
154 | ||
155 | } else { | |
156 | eval { PVE::QemuServer::vm_mon_cmd($vmid, "object-add", 'qom-type' => "memory-backend-ram", id => "mem-$name", props => { size => int($dimm_size*1024*1024) } ) }; | |
157 | } | |
158 | ||
6779f1ac AD |
159 | if (my $err = $@) { |
160 | eval { PVE::QemuServer::qemu_objectdel($vmid, "mem-$name"); }; | |
161 | die $err; | |
162 | } | |
163 | ||
164 | eval { PVE::QemuServer::vm_mon_cmd($vmid, "device_add", driver => "pc-dimm", id => "$name", memdev => "mem-$name", node => $numanode) }; | |
165 | if (my $err = $@) { | |
166 | eval { PVE::QemuServer::qemu_objectdel($vmid, "mem-$name"); }; | |
167 | die $err; | |
168 | } | |
169 | #update conf after each succesful module hotplug | |
170 | $conf->{memory} = $current_size; | |
171 | PVE::QemuConfig->write_config($vmid, $conf); | |
172 | }); | |
173 | ||
174 | } else { | |
175 | ||
176 | foreach_reverse_dimm($conf, $vmid, $value, $sockets, sub { | |
177 | my ($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory) = @_; | |
178 | ||
179 | return if $current_size >= $conf->{memory}; | |
180 | print "try to unplug memory dimm $name\n"; | |
181 | ||
182 | my $retry = 0; | |
183 | while (1) { | |
184 | eval { PVE::QemuServer::qemu_devicedel($vmid, $name) }; | |
185 | sleep 3; | |
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; | |
189 | $retry++; | |
190 | } | |
191 | ||
192 | #update conf after each succesful module unplug | |
193 | $conf->{memory} = $current_size; | |
194 | ||
195 | eval { PVE::QemuServer::qemu_objectdel($vmid, "mem-$name"); }; | |
196 | PVE::QemuConfig->write_config($vmid, $conf); | |
197 | }); | |
198 | } | |
199 | } | |
200 | ||
201 | sub qemu_dimm_list { | |
202 | my ($vmid) = @_; | |
203 | ||
204 | my $dimmarray = PVE::QemuServer::vm_mon_cmd_nocheck($vmid, "query-memory-devices"); | |
205 | my $dimms = {}; | |
206 | ||
207 | foreach my $dimm (@$dimmarray) { | |
208 | ||
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}; | |
214 | } | |
215 | return $dimms; | |
216 | } | |
217 | ||
0567a4d5 AD |
218 | sub config { |
219 | my ($conf, $vmid, $sockets, $cores, $defaults, $hotplug_features, $cmd) = @_; | |
a022e3fd | 220 | |
0567a4d5 AD |
221 | my $memory = $conf->{memory} || $defaults->{memory}; |
222 | my $static_memory = 0; | |
0567a4d5 AD |
223 | |
224 | if ($hotplug_features->{memory}) { | |
7023f3ea | 225 | die "NUMA needs to be enabled for memory hotplug\n" if !$conf->{numa}; |
0567a4d5 | 226 | die "Total memory is bigger than ${MAX_MEM}MB\n" if $memory > $MAX_MEM; |
7023f3ea AD |
227 | my $sockets = 1; |
228 | $sockets = $conf->{sockets} if $conf->{sockets}; | |
229 | ||
0567a4d5 | 230 | $static_memory = $STATICMEM; |
7023f3ea AD |
231 | $static_memory = $static_memory * $sockets if ($conf->{hugepages} && $conf->{hugepages} == 1024); |
232 | ||
0567a4d5 | 233 | die "minimum memory must be ${static_memory}MB\n" if($memory < $static_memory); |
0567a4d5 AD |
234 | push @$cmd, '-m', "size=${static_memory},slots=255,maxmem=${MAX_MEM}M"; |
235 | ||
236 | } else { | |
237 | ||
238 | $static_memory = $memory; | |
239 | push @$cmd, '-m', $static_memory; | |
240 | } | |
241 | ||
7023f3ea AD |
242 | die "numa needs to be enabled to use hugepages" if $conf->{hugepages} && !$conf->{numa}; |
243 | ||
0567a4d5 AD |
244 | if ($conf->{numa}) { |
245 | ||
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$i"}); | |
250 | next if !$numa; | |
251 | # memory | |
252 | die "missing NUMA node$i memory value\n" if !$numa->{memory}; | |
253 | my $numa_memory = $numa->{memory}; | |
254 | $numa_totalmemory += $numa_memory; | |
7023f3ea AD |
255 | |
256 | my $mem_object = print_mem_object($conf, "ram-node$i", $numa_memory); | |
0567a4d5 AD |
257 | |
258 | # cpus | |
259 | my $cpulists = $numa->{cpus}; | |
260 | die "missing NUMA node$i cpus\n" if !defined($cpulists); | |
f612bd67 | 261 | my $cpus = join(',cpus=', map { |
0567a4d5 AD |
262 | my ($start, $end) = @$_; |
263 | defined($end) ? "$start-$end" : $start | |
264 | } @$cpulists); | |
265 | ||
266 | # hostnodes | |
267 | my $hostnodelists = $numa->{hostnodes}; | |
268 | if (defined($hostnodelists)) { | |
ac7b7087 AD |
269 | |
270 | my $hostnodes = print_numa_hostnodes($hostnodelists); | |
0567a4d5 AD |
271 | |
272 | # policy | |
273 | my $policy = $numa->{policy}; | |
274 | die "you need to define a policy for hostnode $hostnodes\n" if !$policy; | |
7023f3ea AD |
275 | $mem_object .= ",host-nodes=$hostnodes,policy=$policy"; |
276 | } else { | |
277 | die "numa hostnodes need to be defined to use hugepages" if $conf->{hugepages}; | |
0567a4d5 AD |
278 | } |
279 | ||
7023f3ea | 280 | push @$cmd, '-object', $mem_object; |
0567a4d5 AD |
281 | push @$cmd, '-numa', "node,nodeid=$i,cpus=$cpus,memdev=ram-node$i"; |
282 | } | |
283 | ||
284 | die "total memory for NUMA nodes must be equal to vm static memory\n" | |
285 | if $numa_totalmemory && $numa_totalmemory != $static_memory; | |
286 | ||
287 | #if no custom tology, we split memory and cores across numa nodes | |
288 | if(!$numa_totalmemory) { | |
289 | ||
7023f3ea | 290 | my $numa_memory = ($static_memory / $sockets); |
0567a4d5 AD |
291 | |
292 | for (my $i = 0; $i < $sockets; $i++) { | |
a2a5ffc9 | 293 | die "host NUMA node$i doesn't exist\n" if ! -d "/sys/devices/system/node/node$i/" && $conf->{hugepages}; |
0567a4d5 AD |
294 | |
295 | my $cpustart = ($cores * $i); | |
296 | my $cpuend = ($cpustart + $cores - 1) if $cores && $cores > 1; | |
297 | my $cpus = $cpustart; | |
298 | $cpus .= "-$cpuend" if $cpuend; | |
299 | ||
7023f3ea AD |
300 | my $mem_object = print_mem_object($conf, "ram-node$i", $numa_memory); |
301 | ||
302 | push @$cmd, '-object', $mem_object; | |
0567a4d5 AD |
303 | push @$cmd, '-numa', "node,nodeid=$i,cpus=$cpus,memdev=ram-node$i"; |
304 | } | |
305 | } | |
306 | } | |
307 | ||
308 | if ($hotplug_features->{memory}) { | |
309 | foreach_dimm($conf, $vmid, $memory, $sockets, sub { | |
310 | my ($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory) = @_; | |
7023f3ea AD |
311 | |
312 | my $mem_object = print_mem_object($conf, "mem-$name", $dimm_size); | |
313 | ||
314 | push @$cmd, "-object" , $mem_object; | |
0567a4d5 AD |
315 | push @$cmd, "-device", "pc-dimm,id=$name,memdev=mem-$name,node=$numanode"; |
316 | ||
317 | #if dimm_memory is not aligned to dimm map | |
318 | if($current_size > $memory) { | |
319 | $conf->{memory} = $current_size; | |
320 | PVE::QemuConfig->write_config($vmid, $conf); | |
321 | } | |
322 | }); | |
323 | } | |
324 | } | |
325 | ||
7023f3ea AD |
326 | sub print_mem_object { |
327 | my ($conf, $id, $size) = @_; | |
328 | ||
329 | if ($conf->{hugepages}) { | |
330 | ||
331 | my $hugepages_size = hugepages_size($conf, $size); | |
332 | my $path = hugepages_mount_path($hugepages_size); | |
333 | ||
334 | return "memory-backend-file,id=$id,size=${size}M,mem-path=$path,share=on,prealloc=yes"; | |
335 | } else { | |
336 | return "memory-backend-ram,id=$id,size=${size}M"; | |
337 | } | |
0567a4d5 | 338 | |
7023f3ea AD |
339 | } |
340 | ||
ac7b7087 AD |
341 | sub print_numa_hostnodes { |
342 | my ($hostnodelists) = @_; | |
343 | ||
344 | my $hostnodes; | |
345 | foreach my $hostnoderange (@$hostnodelists) { | |
346 | my ($start, $end) = @$hostnoderange; | |
347 | $hostnodes .= ',' if $hostnodes; | |
348 | $hostnodes .= $start; | |
349 | $hostnodes .= "-$end" if defined($end); | |
350 | $end //= $start; | |
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/"; | |
353 | } | |
354 | } | |
355 | return $hostnodes; | |
356 | } | |
357 | ||
7023f3ea AD |
358 | sub hugepages_mount { |
359 | ||
360 | my $mountdata = PVE::ProcFSTools::parse_proc_mounts(); | |
361 | ||
362 | foreach my $size (qw(2048 1048576)) { | |
3d8d2e8d | 363 | next if (! -d "/sys/kernel/mm/hugepages/hugepages-${size}kB"); |
7023f3ea AD |
364 | |
365 | my $path = "/run/hugepages/kvm/${size}kB"; | |
366 | ||
367 | my $found = grep { | |
368 | $_->[2] =~ /^hugetlbfs/ && | |
369 | $_->[1] eq $path | |
370 | } @$mountdata; | |
371 | ||
372 | if (!$found) { | |
373 | ||
374 | File::Path::make_path($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"); | |
377 | } | |
378 | } | |
379 | } | |
380 | ||
381 | sub hugepages_mount_path { | |
382 | my ($size) = @_; | |
383 | ||
384 | $size = $size * 1024; | |
385 | return "/run/hugepages/kvm/${size}kB"; | |
386 | ||
387 | } | |
388 | ||
389 | sub hugepages_nr { | |
390 | my ($size, $hugepages_size) = @_; | |
391 | ||
392 | return $size / $hugepages_size; | |
393 | } | |
394 | ||
395 | sub hugepages_size { | |
71aba4ea TL |
396 | my ($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; | |
7023f3ea | 400 | |
71aba4ea TL |
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); | |
7023f3ea | 403 | |
71aba4ea | 404 | if ($conf->{hugepages} eq 'any') { |
7023f3ea | 405 | |
71aba4ea TL |
406 | # try to use 1GB if available && memory size is matching |
407 | if ($page_chunk->(1024) && ($size & 1023) == 0) { | |
7023f3ea | 408 | return 1024; |
71aba4ea | 409 | } elsif ($page_chunk->(2)) { |
7023f3ea | 410 | return 2; |
71aba4ea TL |
411 | } else { |
412 | die "host only supports 1024 GB hugepages, but requested size '$size' is not a multiple of 1024 MB\n" | |
7023f3ea | 413 | } |
71aba4ea | 414 | } else { |
7023f3ea | 415 | |
93981fa7 | 416 | my $hugepagesize = $conf->{hugepages}; |
7023f3ea | 417 | |
93981fa7 TL |
418 | if (!$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"; | |
7023f3ea | 422 | } |
93981fa7 TL |
423 | |
424 | return $hugepagesize | |
71aba4ea | 425 | } |
7023f3ea AD |
426 | } |
427 | ||
428 | sub hugepages_topology { | |
429 | my ($conf) = @_; | |
430 | ||
431 | my $hugepages_topology = {}; | |
432 | ||
433 | return if !$conf->{numa}; | |
434 | ||
435 | my $defaults = PVE::QemuServer::load_defaults(); | |
436 | my $memory = $conf->{memory} || $defaults->{memory}; | |
437 | my $static_memory = 0; | |
438 | my $sockets = 1; | |
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}) ? $conf->{hotplug} : '1'); | |
443 | ||
444 | if ($hotplug_features->{memory}) { | |
445 | $static_memory = $STATICMEM; | |
446 | $static_memory = $static_memory * $sockets if ($conf->{hugepages} && $conf->{hugepages} == 1024); | |
447 | } else { | |
448 | $static_memory = $memory; | |
449 | } | |
450 | ||
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$i"}); | |
455 | next if !$numa; | |
456 | ||
457 | $numa_custom_topology = 1; | |
458 | my $numa_memory = $numa->{memory}; | |
ac7b7087 AD |
459 | my $hostnodelists = $numa->{hostnodes}; |
460 | my $hostnodes = print_numa_hostnodes($hostnodelists); | |
7023f3ea | 461 | |
ac7b7087 | 462 | die "more than 1 hostnode value in numa node is not supported when hugepages are enabled" if $hostnodes !~ m/^(\d)$/; |
7023f3ea | 463 | my $hugepages_size = hugepages_size($conf, $numa_memory); |
ac7b7087 | 464 | $hugepages_topology->{$hugepages_size}->{$hostnodes} += hugepages_nr($numa_memory, $hugepages_size); |
7023f3ea AD |
465 | |
466 | } | |
467 | ||
468 | #if no custom numa tology, we split memory and cores across numa nodes | |
469 | if(!$numa_custom_topology) { | |
470 | ||
471 | my $numa_memory = ($static_memory / $sockets); | |
472 | ||
473 | for (my $i = 0; $i < $sockets; $i++) { | |
474 | ||
475 | my $hugepages_size = hugepages_size($conf, $numa_memory); | |
476 | $hugepages_topology->{$hugepages_size}->{$i} += hugepages_nr($numa_memory, $hugepages_size); | |
477 | } | |
478 | } | |
479 | ||
480 | if ($hotplug_features->{memory}) { | |
aaff69ad WB |
481 | my $numa_hostmap = get_numa_guest_to_host_map($conf); |
482 | ||
7023f3ea AD |
483 | foreach_dimm($conf, undef, $memory, $sockets, sub { |
484 | my ($conf, undef, $name, $dimm_size, $numanode, $current_size, $memory) = @_; | |
485 | ||
aaff69ad WB |
486 | $numanode = $numa_hostmap->{$numanode}; |
487 | ||
7023f3ea AD |
488 | my $hugepages_size = hugepages_size($conf, $dimm_size); |
489 | $hugepages_topology->{$hugepages_size}->{$numanode} += hugepages_nr($dimm_size, $hugepages_size); | |
490 | }); | |
491 | } | |
492 | ||
493 | return $hugepages_topology; | |
494 | } | |
495 | ||
496 | sub hugepages_host_topology { | |
497 | ||
498 | #read host hugepages | |
499 | my $hugepages_host_topology = {}; | |
500 | ||
501 | dir_glob_foreach("/sys/devices/system/node/", 'node(\d+)', sub { | |
502 | my ($nodepath, $numanode) = @_; | |
503 | ||
504 | dir_glob_foreach("/sys/devices/system/node/$nodepath/hugepages/", 'hugepages\-(\d+)kB', sub { | |
505 | my ($hugepages_path, $hugepages_size) = @_; | |
506 | ||
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; | |
510 | }); | |
511 | }); | |
512 | ||
513 | return $hugepages_host_topology; | |
514 | } | |
515 | ||
516 | sub hugepages_allocate { | |
517 | my ($hugepages_topology, $hugepages_host_topology) = @_; | |
518 | ||
519 | #allocate new hupages if needed | |
520 | foreach my $size (sort keys %$hugepages_topology) { | |
521 | ||
522 | my $nodes = $hugepages_topology->{$size}; | |
523 | ||
524 | foreach my $numanode (keys %$nodes) { | |
525 | ||
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"); | |
531 | ||
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"; | |
541 | } | |
542 | } | |
543 | ||
544 | } | |
545 | } | |
546 | ||
547 | } | |
548 | ||
ca0ef6b1 KT |
549 | sub hugepages_default_nr_hugepages { |
550 | my ($size) = @_; | |
551 | ||
552 | my $cmdline = PVE::Tools::file_read_firstline("/proc/cmdline"); | |
553 | my $args = PVE::Tools::split_args($cmdline); | |
554 | ||
555 | my $parsed_size = 2; # default is 2M | |
556 | ||
557 | foreach my $arg (@$args) { | |
558 | if ($arg eq "hugepagesz=2M") { | |
559 | $parsed_size = 2; | |
560 | } elsif ($arg eq "hugepagesz=1G") { | |
561 | $parsed_size = 1024; | |
562 | } elsif ($arg =~ m/^hugepages=(\d+)?$/) { | |
563 | if ($parsed_size == $size) { | |
564 | return $1; | |
565 | } | |
566 | } | |
567 | } | |
568 | ||
569 | return 0; | |
570 | } | |
571 | ||
7023f3ea AD |
572 | sub hugepages_pre_deallocate { |
573 | my ($hugepages_topology) = @_; | |
574 | ||
575 | foreach my $size (sort keys %$hugepages_topology) { | |
576 | ||
577 | my $hugepages_size = $size * 1024; | |
578 | my $path = "/sys/kernel/mm/hugepages/hugepages-${hugepages_size}kB/"; | |
ca0ef6b1 KT |
579 | my $hugepages_nr = hugepages_default_nr_hugepages($size); |
580 | PVE::ProcFSTools::write_proc_entry($path."nr_hugepages", $hugepages_nr); | |
7023f3ea AD |
581 | } |
582 | } | |
583 | ||
584 | sub hugepages_reset { | |
585 | my ($hugepages_topology) = @_; | |
586 | ||
587 | foreach my $size (sort keys %$hugepages_topology) { | |
588 | ||
589 | my $nodes = $hugepages_topology->{$size}; | |
590 | foreach my $numanode (keys %$nodes) { | |
591 | ||
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-${hugepages_size}kB/"; | |
595 | ||
596 | PVE::ProcFSTools::write_proc_entry($path."nr_hugepages", $hugepages_nr); | |
597 | } | |
598 | } | |
599 | } | |
600 | ||
601 | sub hugepages_update_locked { | |
602 | my ($code, @param) = @_; | |
603 | ||
604 | my $timeout = 60; #could be long if a lot of hugepages need to be alocated | |
605 | ||
606 | my $lock_filename = "/var/lock/hugepages.lck"; | |
607 | ||
608 | my $res = lock_file($lock_filename, $timeout, $code, @param); | |
609 | die $@ if $@; | |
610 | ||
611 | return $res; | |
612 | } | |
3f669af2 | 613 | 1; |
a0649da2 | 614 |