xref: /openbmc/qemu/hw/core/numa.c (revision 52924dea)
1 /*
2  * NUMA parameter parsing routines
3  *
4  * Copyright (c) 2014 Fujitsu Ltd.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to deal
8  * in the Software without restriction, including without limitation the rights
9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22  * THE SOFTWARE.
23  */
24 
25 #include "qemu/osdep.h"
26 #include "sysemu/numa.h"
27 #include "exec/cpu-common.h"
28 #include "exec/ramlist.h"
29 #include "qemu/bitmap.h"
30 #include "qemu/error-report.h"
31 #include "qapi/error.h"
32 #include "qapi/opts-visitor.h"
33 #include "qapi/qapi-visit-machine.h"
34 #include "hw/mem/pc-dimm.h"
35 #include "hw/mem/memory-device.h"
36 #include "qemu/option.h"
37 #include "qemu/config-file.h"
38 #include "qemu/cutils.h"
39 
40 QemuOptsList qemu_numa_opts = {
41     .name = "numa",
42     .implied_opt_name = "type",
43     .head = QTAILQ_HEAD_INITIALIZER(qemu_numa_opts.head),
44     .desc = { { 0 } } /* validated with OptsVisitor */
45 };
46 
47 static int have_memdevs = -1;
48 static int max_numa_nodeid; /* Highest specified NUMA node ID, plus one.
49                              * For all nodes, nodeid < max_numa_nodeid
50                              */
51 int nb_numa_nodes;
52 bool have_numa_distance;
53 NodeInfo numa_info[MAX_NODES];
54 
55 
56 static void parse_numa_node(MachineState *ms, NumaNodeOptions *node,
57                             Error **errp)
58 {
59     Error *err = NULL;
60     uint16_t nodenr;
61     uint16List *cpus = NULL;
62     MachineClass *mc = MACHINE_GET_CLASS(ms);
63 
64     if (node->has_nodeid) {
65         nodenr = node->nodeid;
66     } else {
67         nodenr = nb_numa_nodes;
68     }
69 
70     if (nodenr >= MAX_NODES) {
71         error_setg(errp, "Max number of NUMA nodes reached: %"
72                    PRIu16 "", nodenr);
73         return;
74     }
75 
76     if (numa_info[nodenr].present) {
77         error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr);
78         return;
79     }
80 
81     if (!mc->cpu_index_to_instance_props || !mc->get_default_cpu_node_id) {
82         error_setg(errp, "NUMA is not supported by this machine-type");
83         return;
84     }
85     for (cpus = node->cpus; cpus; cpus = cpus->next) {
86         CpuInstanceProperties props;
87         if (cpus->value >= max_cpus) {
88             error_setg(errp,
89                        "CPU index (%" PRIu16 ")"
90                        " should be smaller than maxcpus (%d)",
91                        cpus->value, max_cpus);
92             return;
93         }
94         props = mc->cpu_index_to_instance_props(ms, cpus->value);
95         props.node_id = nodenr;
96         props.has_node_id = true;
97         machine_set_cpu_numa_node(ms, &props, &err);
98         if (err) {
99             error_propagate(errp, err);
100             return;
101         }
102     }
103 
104     if (node->has_mem && node->has_memdev) {
105         error_setg(errp, "cannot specify both mem= and memdev=");
106         return;
107     }
108 
109     if (have_memdevs == -1) {
110         have_memdevs = node->has_memdev;
111     }
112     if (node->has_memdev != have_memdevs) {
113         error_setg(errp, "memdev option must be specified for either "
114                    "all or no nodes");
115         return;
116     }
117 
118     if (node->has_mem) {
119         numa_info[nodenr].node_mem = node->mem;
120     }
121     if (node->has_memdev) {
122         Object *o;
123         o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
124         if (!o) {
125             error_setg(errp, "memdev=%s is ambiguous", node->memdev);
126             return;
127         }
128 
129         object_ref(o);
130         numa_info[nodenr].node_mem = object_property_get_uint(o, "size", NULL);
131         numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
132     }
133     numa_info[nodenr].present = true;
134     max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
135     nb_numa_nodes++;
136 }
137 
138 static void parse_numa_distance(NumaDistOptions *dist, Error **errp)
139 {
140     uint16_t src = dist->src;
141     uint16_t dst = dist->dst;
142     uint8_t val = dist->val;
143 
144     if (src >= MAX_NODES || dst >= MAX_NODES) {
145         error_setg(errp, "Parameter '%s' expects an integer between 0 and %d",
146                    src >= MAX_NODES ? "src" : "dst", MAX_NODES - 1);
147         return;
148     }
149 
150     if (!numa_info[src].present || !numa_info[dst].present) {
151         error_setg(errp, "Source/Destination NUMA node is missing. "
152                    "Please use '-numa node' option to declare it first.");
153         return;
154     }
155 
156     if (val < NUMA_DISTANCE_MIN) {
157         error_setg(errp, "NUMA distance (%" PRIu8 ") is invalid, "
158                    "it shouldn't be less than %d.",
159                    val, NUMA_DISTANCE_MIN);
160         return;
161     }
162 
163     if (src == dst && val != NUMA_DISTANCE_MIN) {
164         error_setg(errp, "Local distance of node %d should be %d.",
165                    src, NUMA_DISTANCE_MIN);
166         return;
167     }
168 
169     numa_info[src].distance[dst] = val;
170     have_numa_distance = true;
171 }
172 
173 void set_numa_options(MachineState *ms, NumaOptions *object, Error **errp)
174 {
175     Error *err = NULL;
176 
177     switch (object->type) {
178     case NUMA_OPTIONS_TYPE_NODE:
179         parse_numa_node(ms, &object->u.node, &err);
180         if (err) {
181             goto end;
182         }
183         break;
184     case NUMA_OPTIONS_TYPE_DIST:
185         parse_numa_distance(&object->u.dist, &err);
186         if (err) {
187             goto end;
188         }
189         break;
190     case NUMA_OPTIONS_TYPE_CPU:
191         if (!object->u.cpu.has_node_id) {
192             error_setg(&err, "Missing mandatory node-id property");
193             goto end;
194         }
195         if (!numa_info[object->u.cpu.node_id].present) {
196             error_setg(&err, "Invalid node-id=%" PRId64 ", NUMA node must be "
197                 "defined with -numa node,nodeid=ID before it's used with "
198                 "-numa cpu,node-id=ID", object->u.cpu.node_id);
199             goto end;
200         }
201 
202         machine_set_cpu_numa_node(ms, qapi_NumaCpuOptions_base(&object->u.cpu),
203                                   &err);
204         break;
205     default:
206         abort();
207     }
208 
209 end:
210     error_propagate(errp, err);
211 }
212 
213 static int parse_numa(void *opaque, QemuOpts *opts, Error **errp)
214 {
215     NumaOptions *object = NULL;
216     MachineState *ms = MACHINE(opaque);
217     Error *err = NULL;
218     Visitor *v = opts_visitor_new(opts);
219 
220     visit_type_NumaOptions(v, NULL, &object, &err);
221     visit_free(v);
222     if (err) {
223         goto end;
224     }
225 
226     /* Fix up legacy suffix-less format */
227     if ((object->type == NUMA_OPTIONS_TYPE_NODE) && object->u.node.has_mem) {
228         const char *mem_str = qemu_opt_get(opts, "mem");
229         qemu_strtosz_MiB(mem_str, NULL, &object->u.node.mem);
230     }
231 
232     set_numa_options(ms, object, &err);
233 
234 end:
235     qapi_free_NumaOptions(object);
236     if (err) {
237         error_propagate(errp, err);
238         return -1;
239     }
240 
241     return 0;
242 }
243 
244 /* If all node pair distances are symmetric, then only distances
245  * in one direction are enough. If there is even one asymmetric
246  * pair, though, then all distances must be provided. The
247  * distance from a node to itself is always NUMA_DISTANCE_MIN,
248  * so providing it is never necessary.
249  */
250 static void validate_numa_distance(void)
251 {
252     int src, dst;
253     bool is_asymmetrical = false;
254 
255     for (src = 0; src < nb_numa_nodes; src++) {
256         for (dst = src; dst < nb_numa_nodes; dst++) {
257             if (numa_info[src].distance[dst] == 0 &&
258                 numa_info[dst].distance[src] == 0) {
259                 if (src != dst) {
260                     error_report("The distance between node %d and %d is "
261                                  "missing, at least one distance value "
262                                  "between each nodes should be provided.",
263                                  src, dst);
264                     exit(EXIT_FAILURE);
265                 }
266             }
267 
268             if (numa_info[src].distance[dst] != 0 &&
269                 numa_info[dst].distance[src] != 0 &&
270                 numa_info[src].distance[dst] !=
271                 numa_info[dst].distance[src]) {
272                 is_asymmetrical = true;
273             }
274         }
275     }
276 
277     if (is_asymmetrical) {
278         for (src = 0; src < nb_numa_nodes; src++) {
279             for (dst = 0; dst < nb_numa_nodes; dst++) {
280                 if (src != dst && numa_info[src].distance[dst] == 0) {
281                     error_report("At least one asymmetrical pair of "
282                             "distances is given, please provide distances "
283                             "for both directions of all node pairs.");
284                     exit(EXIT_FAILURE);
285                 }
286             }
287         }
288     }
289 }
290 
291 static void complete_init_numa_distance(void)
292 {
293     int src, dst;
294 
295     /* Fixup NUMA distance by symmetric policy because if it is an
296      * asymmetric distance table, it should be a complete table and
297      * there would not be any missing distance except local node, which
298      * is verified by validate_numa_distance above.
299      */
300     for (src = 0; src < nb_numa_nodes; src++) {
301         for (dst = 0; dst < nb_numa_nodes; dst++) {
302             if (numa_info[src].distance[dst] == 0) {
303                 if (src == dst) {
304                     numa_info[src].distance[dst] = NUMA_DISTANCE_MIN;
305                 } else {
306                     numa_info[src].distance[dst] = numa_info[dst].distance[src];
307                 }
308             }
309         }
310     }
311 }
312 
313 void numa_legacy_auto_assign_ram(MachineClass *mc, NodeInfo *nodes,
314                                  int nb_nodes, ram_addr_t size)
315 {
316     int i;
317     uint64_t usedmem = 0;
318 
319     /* Align each node according to the alignment
320      * requirements of the machine class
321      */
322 
323     for (i = 0; i < nb_nodes - 1; i++) {
324         nodes[i].node_mem = (size / nb_nodes) &
325                             ~((1 << mc->numa_mem_align_shift) - 1);
326         usedmem += nodes[i].node_mem;
327     }
328     nodes[i].node_mem = size - usedmem;
329 }
330 
331 void numa_default_auto_assign_ram(MachineClass *mc, NodeInfo *nodes,
332                                   int nb_nodes, ram_addr_t size)
333 {
334     int i;
335     uint64_t usedmem = 0, node_mem;
336     uint64_t granularity = size / nb_nodes;
337     uint64_t propagate = 0;
338 
339     for (i = 0; i < nb_nodes - 1; i++) {
340         node_mem = (granularity + propagate) &
341                    ~((1 << mc->numa_mem_align_shift) - 1);
342         propagate = granularity + propagate - node_mem;
343         nodes[i].node_mem = node_mem;
344         usedmem += node_mem;
345     }
346     nodes[i].node_mem = size - usedmem;
347 }
348 
349 void numa_complete_configuration(MachineState *ms)
350 {
351     int i;
352     MachineClass *mc = MACHINE_GET_CLASS(ms);
353 
354     /*
355      * If memory hotplug is enabled (slots > 0) but without '-numa'
356      * options explicitly on CLI, guestes will break.
357      *
358      *   Windows: won't enable memory hotplug without SRAT table at all
359      *
360      *   Linux: if QEMU is started with initial memory all below 4Gb
361      *   and no SRAT table present, guest kernel will use nommu DMA ops,
362      *   which breaks 32bit hw drivers when memory is hotplugged and
363      *   guest tries to use it with that drivers.
364      *
365      * Enable NUMA implicitly by adding a new NUMA node automatically.
366      */
367     if (ms->ram_slots > 0 && nb_numa_nodes == 0 &&
368         mc->auto_enable_numa_with_memhp) {
369             NumaNodeOptions node = { };
370             parse_numa_node(ms, &node, &error_abort);
371     }
372 
373     assert(max_numa_nodeid <= MAX_NODES);
374 
375     /* No support for sparse NUMA node IDs yet: */
376     for (i = max_numa_nodeid - 1; i >= 0; i--) {
377         /* Report large node IDs first, to make mistakes easier to spot */
378         if (!numa_info[i].present) {
379             error_report("numa: Node ID missing: %d", i);
380             exit(1);
381         }
382     }
383 
384     /* This must be always true if all nodes are present: */
385     assert(nb_numa_nodes == max_numa_nodeid);
386 
387     if (nb_numa_nodes > 0) {
388         uint64_t numa_total;
389 
390         if (nb_numa_nodes > MAX_NODES) {
391             nb_numa_nodes = MAX_NODES;
392         }
393 
394         /* If no memory size is given for any node, assume the default case
395          * and distribute the available memory equally across all nodes
396          */
397         for (i = 0; i < nb_numa_nodes; i++) {
398             if (numa_info[i].node_mem != 0) {
399                 break;
400             }
401         }
402         if (i == nb_numa_nodes) {
403             assert(mc->numa_auto_assign_ram);
404             mc->numa_auto_assign_ram(mc, numa_info, nb_numa_nodes, ram_size);
405         }
406 
407         numa_total = 0;
408         for (i = 0; i < nb_numa_nodes; i++) {
409             numa_total += numa_info[i].node_mem;
410         }
411         if (numa_total != ram_size) {
412             error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
413                          " should equal RAM size (0x" RAM_ADDR_FMT ")",
414                          numa_total, ram_size);
415             exit(1);
416         }
417 
418         /* QEMU needs at least all unique node pair distances to build
419          * the whole NUMA distance table. QEMU treats the distance table
420          * as symmetric by default, i.e. distance A->B == distance B->A.
421          * Thus, QEMU is able to complete the distance table
422          * initialization even though only distance A->B is provided and
423          * distance B->A is not. QEMU knows the distance of a node to
424          * itself is always 10, so A->A distances may be omitted. When
425          * the distances of two nodes of a pair differ, i.e. distance
426          * A->B != distance B->A, then that means the distance table is
427          * asymmetric. In this case, the distances for both directions
428          * of all node pairs are required.
429          */
430         if (have_numa_distance) {
431             /* Validate enough NUMA distance information was provided. */
432             validate_numa_distance();
433 
434             /* Validation succeeded, now fill in any missing distances. */
435             complete_init_numa_distance();
436         }
437     }
438 }
439 
440 void parse_numa_opts(MachineState *ms)
441 {
442     qemu_opts_foreach(qemu_find_opts("numa"), parse_numa, ms, &error_fatal);
443 }
444 
445 void numa_cpu_pre_plug(const CPUArchId *slot, DeviceState *dev, Error **errp)
446 {
447     int node_id = object_property_get_int(OBJECT(dev), "node-id", &error_abort);
448 
449     if (node_id == CPU_UNSET_NUMA_NODE_ID) {
450         /* due to bug in libvirt, it doesn't pass node-id from props on
451          * device_add as expected, so we have to fix it up here */
452         if (slot->props.has_node_id) {
453             object_property_set_int(OBJECT(dev), slot->props.node_id,
454                                     "node-id", errp);
455         }
456     } else if (node_id != slot->props.node_id) {
457         error_setg(errp, "invalid node-id, must be %"PRId64,
458                    slot->props.node_id);
459     }
460 }
461 
462 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
463                                            const char *name,
464                                            uint64_t ram_size)
465 {
466     if (mem_path) {
467 #ifdef __linux__
468         Error *err = NULL;
469         memory_region_init_ram_from_file(mr, owner, name, ram_size, 0, 0,
470                                          mem_path, &err);
471         if (err) {
472             error_report_err(err);
473             if (mem_prealloc) {
474                 exit(1);
475             }
476             error_report("falling back to regular RAM allocation.");
477 
478             /* Legacy behavior: if allocation failed, fall back to
479              * regular RAM allocation.
480              */
481             mem_path = NULL;
482             memory_region_init_ram_nomigrate(mr, owner, name, ram_size, &error_fatal);
483         }
484 #else
485         fprintf(stderr, "-mem-path not supported on this host\n");
486         exit(1);
487 #endif
488     } else {
489         memory_region_init_ram_nomigrate(mr, owner, name, ram_size, &error_fatal);
490     }
491     vmstate_register_ram_global(mr);
492 }
493 
494 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
495                                           const char *name,
496                                           uint64_t ram_size)
497 {
498     uint64_t addr = 0;
499     int i;
500 
501     if (nb_numa_nodes == 0 || !have_memdevs) {
502         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
503         return;
504     }
505 
506     memory_region_init(mr, owner, name, ram_size);
507     for (i = 0; i < nb_numa_nodes; i++) {
508         uint64_t size = numa_info[i].node_mem;
509         HostMemoryBackend *backend = numa_info[i].node_memdev;
510         if (!backend) {
511             continue;
512         }
513         MemoryRegion *seg = host_memory_backend_get_memory(backend);
514 
515         if (memory_region_is_mapped(seg)) {
516             char *path = object_get_canonical_path_component(OBJECT(backend));
517             error_report("memory backend %s is used multiple times. Each "
518                          "-numa option must use a different memdev value.",
519                          path);
520             g_free(path);
521             exit(1);
522         }
523 
524         host_memory_backend_set_mapped(backend, true);
525         memory_region_add_subregion(mr, addr, seg);
526         vmstate_register_ram_global(seg);
527         addr += size;
528     }
529 }
530 
531 static void numa_stat_memory_devices(NumaNodeMem node_mem[])
532 {
533     MemoryDeviceInfoList *info_list = qmp_memory_device_list();
534     MemoryDeviceInfoList *info;
535     PCDIMMDeviceInfo     *pcdimm_info;
536 
537     for (info = info_list; info; info = info->next) {
538         MemoryDeviceInfo *value = info->value;
539 
540         if (value) {
541             switch (value->type) {
542             case MEMORY_DEVICE_INFO_KIND_DIMM:
543                 pcdimm_info = value->u.dimm.data;
544                 break;
545 
546             case MEMORY_DEVICE_INFO_KIND_NVDIMM:
547                 pcdimm_info = value->u.nvdimm.data;
548                 break;
549 
550             default:
551                 pcdimm_info = NULL;
552                 break;
553             }
554 
555             if (pcdimm_info) {
556                 node_mem[pcdimm_info->node].node_mem += pcdimm_info->size;
557                 node_mem[pcdimm_info->node].node_plugged_mem +=
558                     pcdimm_info->size;
559             }
560         }
561     }
562     qapi_free_MemoryDeviceInfoList(info_list);
563 }
564 
565 void query_numa_node_mem(NumaNodeMem node_mem[])
566 {
567     int i;
568 
569     if (nb_numa_nodes <= 0) {
570         return;
571     }
572 
573     numa_stat_memory_devices(node_mem);
574     for (i = 0; i < nb_numa_nodes; i++) {
575         node_mem[i].node_mem += numa_info[i].node_mem;
576     }
577 }
578 
579 void ram_block_notifier_add(RAMBlockNotifier *n)
580 {
581     QLIST_INSERT_HEAD(&ram_list.ramblock_notifiers, n, next);
582 }
583 
584 void ram_block_notifier_remove(RAMBlockNotifier *n)
585 {
586     QLIST_REMOVE(n, next);
587 }
588 
589 void ram_block_notify_add(void *host, size_t size)
590 {
591     RAMBlockNotifier *notifier;
592 
593     QLIST_FOREACH(notifier, &ram_list.ramblock_notifiers, next) {
594         notifier->ram_block_added(notifier, host, size);
595     }
596 }
597 
598 void ram_block_notify_remove(void *host, size_t size)
599 {
600     RAMBlockNotifier *notifier;
601 
602     QLIST_FOREACH(notifier, &ram_list.ramblock_notifiers, next) {
603         notifier->ram_block_removed(notifier, host, size);
604     }
605 }
606