~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/arch/powerpc/mm/numa.c

Version: ~ [ linux-5.4-rc3 ] ~ [ linux-5.3.6 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.79 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.149 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.196 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.196 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.75 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * pSeries NUMA support
  3  *
  4  * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
  5  *
  6  * This program is free software; you can redistribute it and/or
  7  * modify it under the terms of the GNU General Public License
  8  * as published by the Free Software Foundation; either version
  9  * 2 of the License, or (at your option) any later version.
 10  */
 11 #define pr_fmt(fmt) "numa: " fmt
 12 
 13 #include <linux/threads.h>
 14 #include <linux/bootmem.h>
 15 #include <linux/init.h>
 16 #include <linux/mm.h>
 17 #include <linux/mmzone.h>
 18 #include <linux/export.h>
 19 #include <linux/nodemask.h>
 20 #include <linux/cpu.h>
 21 #include <linux/notifier.h>
 22 #include <linux/memblock.h>
 23 #include <linux/of.h>
 24 #include <linux/pfn.h>
 25 #include <linux/cpuset.h>
 26 #include <linux/node.h>
 27 #include <linux/stop_machine.h>
 28 #include <linux/proc_fs.h>
 29 #include <linux/seq_file.h>
 30 #include <linux/uaccess.h>
 31 #include <linux/slab.h>
 32 #include <asm/cputhreads.h>
 33 #include <asm/sparsemem.h>
 34 #include <asm/prom.h>
 35 #include <asm/smp.h>
 36 #include <asm/cputhreads.h>
 37 #include <asm/topology.h>
 38 #include <asm/firmware.h>
 39 #include <asm/paca.h>
 40 #include <asm/hvcall.h>
 41 #include <asm/setup.h>
 42 #include <asm/vdso.h>
 43 
 44 static int numa_enabled = 1;
 45 
 46 static char *cmdline __initdata;
 47 
 48 static int numa_debug;
 49 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
 50 
 51 int numa_cpu_lookup_table[NR_CPUS];
 52 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
 53 struct pglist_data *node_data[MAX_NUMNODES];
 54 
 55 EXPORT_SYMBOL(numa_cpu_lookup_table);
 56 EXPORT_SYMBOL(node_to_cpumask_map);
 57 EXPORT_SYMBOL(node_data);
 58 
 59 static int min_common_depth;
 60 static int n_mem_addr_cells, n_mem_size_cells;
 61 static int form1_affinity;
 62 
 63 #define MAX_DISTANCE_REF_POINTS 4
 64 static int distance_ref_points_depth;
 65 static const __be32 *distance_ref_points;
 66 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
 67 
 68 /*
 69  * Allocate node_to_cpumask_map based on number of available nodes
 70  * Requires node_possible_map to be valid.
 71  *
 72  * Note: cpumask_of_node() is not valid until after this is done.
 73  */
 74 static void __init setup_node_to_cpumask_map(void)
 75 {
 76         unsigned int node;
 77 
 78         /* setup nr_node_ids if not done yet */
 79         if (nr_node_ids == MAX_NUMNODES)
 80                 setup_nr_node_ids();
 81 
 82         /* allocate the map */
 83         for_each_node(node)
 84                 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
 85 
 86         /* cpumask_of_node() will now work */
 87         dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
 88 }
 89 
 90 static int __init fake_numa_create_new_node(unsigned long end_pfn,
 91                                                 unsigned int *nid)
 92 {
 93         unsigned long long mem;
 94         char *p = cmdline;
 95         static unsigned int fake_nid;
 96         static unsigned long long curr_boundary;
 97 
 98         /*
 99          * Modify node id, iff we started creating NUMA nodes
100          * We want to continue from where we left of the last time
101          */
102         if (fake_nid)
103                 *nid = fake_nid;
104         /*
105          * In case there are no more arguments to parse, the
106          * node_id should be the same as the last fake node id
107          * (we've handled this above).
108          */
109         if (!p)
110                 return 0;
111 
112         mem = memparse(p, &p);
113         if (!mem)
114                 return 0;
115 
116         if (mem < curr_boundary)
117                 return 0;
118 
119         curr_boundary = mem;
120 
121         if ((end_pfn << PAGE_SHIFT) > mem) {
122                 /*
123                  * Skip commas and spaces
124                  */
125                 while (*p == ',' || *p == ' ' || *p == '\t')
126                         p++;
127 
128                 cmdline = p;
129                 fake_nid++;
130                 *nid = fake_nid;
131                 dbg("created new fake_node with id %d\n", fake_nid);
132                 return 1;
133         }
134         return 0;
135 }
136 
137 static void reset_numa_cpu_lookup_table(void)
138 {
139         unsigned int cpu;
140 
141         for_each_possible_cpu(cpu)
142                 numa_cpu_lookup_table[cpu] = -1;
143 }
144 
145 static void update_numa_cpu_lookup_table(unsigned int cpu, int node)
146 {
147         numa_cpu_lookup_table[cpu] = node;
148 }
149 
150 static void map_cpu_to_node(int cpu, int node)
151 {
152         update_numa_cpu_lookup_table(cpu, node);
153 
154         dbg("adding cpu %d to node %d\n", cpu, node);
155 
156         if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
157                 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
158 }
159 
160 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
161 static void unmap_cpu_from_node(unsigned long cpu)
162 {
163         int node = numa_cpu_lookup_table[cpu];
164 
165         dbg("removing cpu %lu from node %d\n", cpu, node);
166 
167         if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
168                 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
169         } else {
170                 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
171                        cpu, node);
172         }
173 }
174 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
175 
176 /* must hold reference to node during call */
177 static const __be32 *of_get_associativity(struct device_node *dev)
178 {
179         return of_get_property(dev, "ibm,associativity", NULL);
180 }
181 
182 /*
183  * Returns the property linux,drconf-usable-memory if
184  * it exists (the property exists only in kexec/kdump kernels,
185  * added by kexec-tools)
186  */
187 static const __be32 *of_get_usable_memory(struct device_node *memory)
188 {
189         const __be32 *prop;
190         u32 len;
191         prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
192         if (!prop || len < sizeof(unsigned int))
193                 return NULL;
194         return prop;
195 }
196 
197 int __node_distance(int a, int b)
198 {
199         int i;
200         int distance = LOCAL_DISTANCE;
201 
202         if (!form1_affinity)
203                 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
204 
205         for (i = 0; i < distance_ref_points_depth; i++) {
206                 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
207                         break;
208 
209                 /* Double the distance for each NUMA level */
210                 distance *= 2;
211         }
212 
213         return distance;
214 }
215 EXPORT_SYMBOL(__node_distance);
216 
217 static void initialize_distance_lookup_table(int nid,
218                 const __be32 *associativity)
219 {
220         int i;
221 
222         if (!form1_affinity)
223                 return;
224 
225         for (i = 0; i < distance_ref_points_depth; i++) {
226                 const __be32 *entry;
227 
228                 entry = &associativity[be32_to_cpu(distance_ref_points[i]) - 1];
229                 distance_lookup_table[nid][i] = of_read_number(entry, 1);
230         }
231 }
232 
233 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
234  * info is found.
235  */
236 static int associativity_to_nid(const __be32 *associativity)
237 {
238         int nid = -1;
239 
240         if (min_common_depth == -1)
241                 goto out;
242 
243         if (of_read_number(associativity, 1) >= min_common_depth)
244                 nid = of_read_number(&associativity[min_common_depth], 1);
245 
246         /* POWER4 LPAR uses 0xffff as invalid node */
247         if (nid == 0xffff || nid >= MAX_NUMNODES)
248                 nid = -1;
249 
250         if (nid > 0 &&
251                 of_read_number(associativity, 1) >= distance_ref_points_depth) {
252                 /*
253                  * Skip the length field and send start of associativity array
254                  */
255                 initialize_distance_lookup_table(nid, associativity + 1);
256         }
257 
258 out:
259         return nid;
260 }
261 
262 /* Returns the nid associated with the given device tree node,
263  * or -1 if not found.
264  */
265 static int of_node_to_nid_single(struct device_node *device)
266 {
267         int nid = -1;
268         const __be32 *tmp;
269 
270         tmp = of_get_associativity(device);
271         if (tmp)
272                 nid = associativity_to_nid(tmp);
273         return nid;
274 }
275 
276 /* Walk the device tree upwards, looking for an associativity id */
277 int of_node_to_nid(struct device_node *device)
278 {
279         int nid = -1;
280 
281         of_node_get(device);
282         while (device) {
283                 nid = of_node_to_nid_single(device);
284                 if (nid != -1)
285                         break;
286 
287                 device = of_get_next_parent(device);
288         }
289         of_node_put(device);
290 
291         return nid;
292 }
293 EXPORT_SYMBOL_GPL(of_node_to_nid);
294 
295 static int __init find_min_common_depth(void)
296 {
297         int depth;
298         struct device_node *root;
299 
300         if (firmware_has_feature(FW_FEATURE_OPAL))
301                 root = of_find_node_by_path("/ibm,opal");
302         else
303                 root = of_find_node_by_path("/rtas");
304         if (!root)
305                 root = of_find_node_by_path("/");
306 
307         /*
308          * This property is a set of 32-bit integers, each representing
309          * an index into the ibm,associativity nodes.
310          *
311          * With form 0 affinity the first integer is for an SMP configuration
312          * (should be all 0's) and the second is for a normal NUMA
313          * configuration. We have only one level of NUMA.
314          *
315          * With form 1 affinity the first integer is the most significant
316          * NUMA boundary and the following are progressively less significant
317          * boundaries. There can be more than one level of NUMA.
318          */
319         distance_ref_points = of_get_property(root,
320                                         "ibm,associativity-reference-points",
321                                         &distance_ref_points_depth);
322 
323         if (!distance_ref_points) {
324                 dbg("NUMA: ibm,associativity-reference-points not found.\n");
325                 goto err;
326         }
327 
328         distance_ref_points_depth /= sizeof(int);
329 
330         if (firmware_has_feature(FW_FEATURE_OPAL) ||
331             firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
332                 dbg("Using form 1 affinity\n");
333                 form1_affinity = 1;
334         }
335 
336         if (form1_affinity) {
337                 depth = of_read_number(distance_ref_points, 1);
338         } else {
339                 if (distance_ref_points_depth < 2) {
340                         printk(KERN_WARNING "NUMA: "
341                                 "short ibm,associativity-reference-points\n");
342                         goto err;
343                 }
344 
345                 depth = of_read_number(&distance_ref_points[1], 1);
346         }
347 
348         /*
349          * Warn and cap if the hardware supports more than
350          * MAX_DISTANCE_REF_POINTS domains.
351          */
352         if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
353                 printk(KERN_WARNING "NUMA: distance array capped at "
354                         "%d entries\n", MAX_DISTANCE_REF_POINTS);
355                 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
356         }
357 
358         of_node_put(root);
359         return depth;
360 
361 err:
362         of_node_put(root);
363         return -1;
364 }
365 
366 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
367 {
368         struct device_node *memory = NULL;
369 
370         memory = of_find_node_by_type(memory, "memory");
371         if (!memory)
372                 panic("numa.c: No memory nodes found!");
373 
374         *n_addr_cells = of_n_addr_cells(memory);
375         *n_size_cells = of_n_size_cells(memory);
376         of_node_put(memory);
377 }
378 
379 static unsigned long read_n_cells(int n, const __be32 **buf)
380 {
381         unsigned long result = 0;
382 
383         while (n--) {
384                 result = (result << 32) | of_read_number(*buf, 1);
385                 (*buf)++;
386         }
387         return result;
388 }
389 
390 /*
391  * Read the next memblock list entry from the ibm,dynamic-memory property
392  * and return the information in the provided of_drconf_cell structure.
393  */
394 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
395 {
396         const __be32 *cp;
397 
398         drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
399 
400         cp = *cellp;
401         drmem->drc_index = of_read_number(cp, 1);
402         drmem->reserved = of_read_number(&cp[1], 1);
403         drmem->aa_index = of_read_number(&cp[2], 1);
404         drmem->flags = of_read_number(&cp[3], 1);
405 
406         *cellp = cp + 4;
407 }
408 
409 /*
410  * Retrieve and validate the ibm,dynamic-memory property of the device tree.
411  *
412  * The layout of the ibm,dynamic-memory property is a number N of memblock
413  * list entries followed by N memblock list entries.  Each memblock list entry
414  * contains information as laid out in the of_drconf_cell struct above.
415  */
416 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
417 {
418         const __be32 *prop;
419         u32 len, entries;
420 
421         prop = of_get_property(memory, "ibm,dynamic-memory", &len);
422         if (!prop || len < sizeof(unsigned int))
423                 return 0;
424 
425         entries = of_read_number(prop++, 1);
426 
427         /* Now that we know the number of entries, revalidate the size
428          * of the property read in to ensure we have everything
429          */
430         if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
431                 return 0;
432 
433         *dm = prop;
434         return entries;
435 }
436 
437 /*
438  * Retrieve and validate the ibm,lmb-size property for drconf memory
439  * from the device tree.
440  */
441 static u64 of_get_lmb_size(struct device_node *memory)
442 {
443         const __be32 *prop;
444         u32 len;
445 
446         prop = of_get_property(memory, "ibm,lmb-size", &len);
447         if (!prop || len < sizeof(unsigned int))
448                 return 0;
449 
450         return read_n_cells(n_mem_size_cells, &prop);
451 }
452 
453 struct assoc_arrays {
454         u32     n_arrays;
455         u32     array_sz;
456         const __be32 *arrays;
457 };
458 
459 /*
460  * Retrieve and validate the list of associativity arrays for drconf
461  * memory from the ibm,associativity-lookup-arrays property of the
462  * device tree..
463  *
464  * The layout of the ibm,associativity-lookup-arrays property is a number N
465  * indicating the number of associativity arrays, followed by a number M
466  * indicating the size of each associativity array, followed by a list
467  * of N associativity arrays.
468  */
469 static int of_get_assoc_arrays(struct device_node *memory,
470                                struct assoc_arrays *aa)
471 {
472         const __be32 *prop;
473         u32 len;
474 
475         prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
476         if (!prop || len < 2 * sizeof(unsigned int))
477                 return -1;
478 
479         aa->n_arrays = of_read_number(prop++, 1);
480         aa->array_sz = of_read_number(prop++, 1);
481 
482         /* Now that we know the number of arrays and size of each array,
483          * revalidate the size of the property read in.
484          */
485         if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
486                 return -1;
487 
488         aa->arrays = prop;
489         return 0;
490 }
491 
492 /*
493  * This is like of_node_to_nid_single() for memory represented in the
494  * ibm,dynamic-reconfiguration-memory node.
495  */
496 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
497                                    struct assoc_arrays *aa)
498 {
499         int default_nid = 0;
500         int nid = default_nid;
501         int index;
502 
503         if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
504             !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
505             drmem->aa_index < aa->n_arrays) {
506                 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
507                 nid = of_read_number(&aa->arrays[index], 1);
508 
509                 if (nid == 0xffff || nid >= MAX_NUMNODES)
510                         nid = default_nid;
511 
512                 if (nid > 0) {
513                         index = drmem->aa_index * aa->array_sz;
514                         initialize_distance_lookup_table(nid,
515                                                         &aa->arrays[index]);
516                 }
517         }
518 
519         return nid;
520 }
521 
522 /*
523  * Figure out to which domain a cpu belongs and stick it there.
524  * Return the id of the domain used.
525  */
526 static int numa_setup_cpu(unsigned long lcpu)
527 {
528         int nid = -1;
529         struct device_node *cpu;
530 
531         /*
532          * If a valid cpu-to-node mapping is already available, use it
533          * directly instead of querying the firmware, since it represents
534          * the most recent mapping notified to us by the platform (eg: VPHN).
535          */
536         if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
537                 map_cpu_to_node(lcpu, nid);
538                 return nid;
539         }
540 
541         cpu = of_get_cpu_node(lcpu, NULL);
542 
543         if (!cpu) {
544                 WARN_ON(1);
545                 if (cpu_present(lcpu))
546                         goto out_present;
547                 else
548                         goto out;
549         }
550 
551         nid = of_node_to_nid_single(cpu);
552 
553 out_present:
554         if (nid < 0 || !node_online(nid))
555                 nid = first_online_node;
556 
557         map_cpu_to_node(lcpu, nid);
558         of_node_put(cpu);
559 out:
560         return nid;
561 }
562 
563 static void verify_cpu_node_mapping(int cpu, int node)
564 {
565         int base, sibling, i;
566 
567         /* Verify that all the threads in the core belong to the same node */
568         base = cpu_first_thread_sibling(cpu);
569 
570         for (i = 0; i < threads_per_core; i++) {
571                 sibling = base + i;
572 
573                 if (sibling == cpu || cpu_is_offline(sibling))
574                         continue;
575 
576                 if (cpu_to_node(sibling) != node) {
577                         WARN(1, "CPU thread siblings %d and %d don't belong"
578                                 " to the same node!\n", cpu, sibling);
579                         break;
580                 }
581         }
582 }
583 
584 /* Must run before sched domains notifier. */
585 static int ppc_numa_cpu_prepare(unsigned int cpu)
586 {
587         int nid;
588 
589         nid = numa_setup_cpu(cpu);
590         verify_cpu_node_mapping(cpu, nid);
591         return 0;
592 }
593 
594 static int ppc_numa_cpu_dead(unsigned int cpu)
595 {
596 #ifdef CONFIG_HOTPLUG_CPU
597         unmap_cpu_from_node(cpu);
598 #endif
599         return 0;
600 }
601 
602 /*
603  * Check and possibly modify a memory region to enforce the memory limit.
604  *
605  * Returns the size the region should have to enforce the memory limit.
606  * This will either be the original value of size, a truncated value,
607  * or zero. If the returned value of size is 0 the region should be
608  * discarded as it lies wholly above the memory limit.
609  */
610 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
611                                                       unsigned long size)
612 {
613         /*
614          * We use memblock_end_of_DRAM() in here instead of memory_limit because
615          * we've already adjusted it for the limit and it takes care of
616          * having memory holes below the limit.  Also, in the case of
617          * iommu_is_off, memory_limit is not set but is implicitly enforced.
618          */
619 
620         if (start + size <= memblock_end_of_DRAM())
621                 return size;
622 
623         if (start >= memblock_end_of_DRAM())
624                 return 0;
625 
626         return memblock_end_of_DRAM() - start;
627 }
628 
629 /*
630  * Reads the counter for a given entry in
631  * linux,drconf-usable-memory property
632  */
633 static inline int __init read_usm_ranges(const __be32 **usm)
634 {
635         /*
636          * For each lmb in ibm,dynamic-memory a corresponding
637          * entry in linux,drconf-usable-memory property contains
638          * a counter followed by that many (base, size) duple.
639          * read the counter from linux,drconf-usable-memory
640          */
641         return read_n_cells(n_mem_size_cells, usm);
642 }
643 
644 /*
645  * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
646  * node.  This assumes n_mem_{addr,size}_cells have been set.
647  */
648 static void __init parse_drconf_memory(struct device_node *memory)
649 {
650         const __be32 *uninitialized_var(dm), *usm;
651         unsigned int n, rc, ranges, is_kexec_kdump = 0;
652         unsigned long lmb_size, base, size, sz;
653         int nid;
654         struct assoc_arrays aa = { .arrays = NULL };
655 
656         n = of_get_drconf_memory(memory, &dm);
657         if (!n)
658                 return;
659 
660         lmb_size = of_get_lmb_size(memory);
661         if (!lmb_size)
662                 return;
663 
664         rc = of_get_assoc_arrays(memory, &aa);
665         if (rc)
666                 return;
667 
668         /* check if this is a kexec/kdump kernel */
669         usm = of_get_usable_memory(memory);
670         if (usm != NULL)
671                 is_kexec_kdump = 1;
672 
673         for (; n != 0; --n) {
674                 struct of_drconf_cell drmem;
675 
676                 read_drconf_cell(&drmem, &dm);
677 
678                 /* skip this block if the reserved bit is set in flags (0x80)
679                    or if the block is not assigned to this partition (0x8) */
680                 if ((drmem.flags & DRCONF_MEM_RESERVED)
681                     || !(drmem.flags & DRCONF_MEM_ASSIGNED))
682                         continue;
683 
684                 base = drmem.base_addr;
685                 size = lmb_size;
686                 ranges = 1;
687 
688                 if (is_kexec_kdump) {
689                         ranges = read_usm_ranges(&usm);
690                         if (!ranges) /* there are no (base, size) duple */
691                                 continue;
692                 }
693                 do {
694                         if (is_kexec_kdump) {
695                                 base = read_n_cells(n_mem_addr_cells, &usm);
696                                 size = read_n_cells(n_mem_size_cells, &usm);
697                         }
698                         nid = of_drconf_to_nid_single(&drmem, &aa);
699                         fake_numa_create_new_node(
700                                 ((base + size) >> PAGE_SHIFT),
701                                            &nid);
702                         node_set_online(nid);
703                         sz = numa_enforce_memory_limit(base, size);
704                         if (sz)
705                                 memblock_set_node(base, sz,
706                                                   &memblock.memory, nid);
707                 } while (--ranges);
708         }
709 }
710 
711 static int __init parse_numa_properties(void)
712 {
713         struct device_node *memory;
714         int default_nid = 0;
715         unsigned long i;
716 
717         if (numa_enabled == 0) {
718                 printk(KERN_WARNING "NUMA disabled by user\n");
719                 return -1;
720         }
721 
722         min_common_depth = find_min_common_depth();
723 
724         if (min_common_depth < 0)
725                 return min_common_depth;
726 
727         dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
728 
729         /*
730          * Even though we connect cpus to numa domains later in SMP
731          * init, we need to know the node ids now. This is because
732          * each node to be onlined must have NODE_DATA etc backing it.
733          */
734         for_each_present_cpu(i) {
735                 struct device_node *cpu;
736                 int nid;
737 
738                 cpu = of_get_cpu_node(i, NULL);
739                 BUG_ON(!cpu);
740                 nid = of_node_to_nid_single(cpu);
741                 of_node_put(cpu);
742 
743                 /*
744                  * Don't fall back to default_nid yet -- we will plug
745                  * cpus into nodes once the memory scan has discovered
746                  * the topology.
747                  */
748                 if (nid < 0)
749                         continue;
750                 node_set_online(nid);
751         }
752 
753         get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
754 
755         for_each_node_by_type(memory, "memory") {
756                 unsigned long start;
757                 unsigned long size;
758                 int nid;
759                 int ranges;
760                 const __be32 *memcell_buf;
761                 unsigned int len;
762 
763                 memcell_buf = of_get_property(memory,
764                         "linux,usable-memory", &len);
765                 if (!memcell_buf || len <= 0)
766                         memcell_buf = of_get_property(memory, "reg", &len);
767                 if (!memcell_buf || len <= 0)
768                         continue;
769 
770                 /* ranges in cell */
771                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
772 new_range:
773                 /* these are order-sensitive, and modify the buffer pointer */
774                 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
775                 size = read_n_cells(n_mem_size_cells, &memcell_buf);
776 
777                 /*
778                  * Assumption: either all memory nodes or none will
779                  * have associativity properties.  If none, then
780                  * everything goes to default_nid.
781                  */
782                 nid = of_node_to_nid_single(memory);
783                 if (nid < 0)
784                         nid = default_nid;
785 
786                 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
787                 node_set_online(nid);
788 
789                 if (!(size = numa_enforce_memory_limit(start, size))) {
790                         if (--ranges)
791                                 goto new_range;
792                         else
793                                 continue;
794                 }
795 
796                 memblock_set_node(start, size, &memblock.memory, nid);
797 
798                 if (--ranges)
799                         goto new_range;
800         }
801 
802         /*
803          * Now do the same thing for each MEMBLOCK listed in the
804          * ibm,dynamic-memory property in the
805          * ibm,dynamic-reconfiguration-memory node.
806          */
807         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
808         if (memory)
809                 parse_drconf_memory(memory);
810 
811         return 0;
812 }
813 
814 static void __init setup_nonnuma(void)
815 {
816         unsigned long top_of_ram = memblock_end_of_DRAM();
817         unsigned long total_ram = memblock_phys_mem_size();
818         unsigned long start_pfn, end_pfn;
819         unsigned int nid = 0;
820         struct memblock_region *reg;
821 
822         printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
823                top_of_ram, total_ram);
824         printk(KERN_DEBUG "Memory hole size: %ldMB\n",
825                (top_of_ram - total_ram) >> 20);
826 
827         for_each_memblock(memory, reg) {
828                 start_pfn = memblock_region_memory_base_pfn(reg);
829                 end_pfn = memblock_region_memory_end_pfn(reg);
830 
831                 fake_numa_create_new_node(end_pfn, &nid);
832                 memblock_set_node(PFN_PHYS(start_pfn),
833                                   PFN_PHYS(end_pfn - start_pfn),
834                                   &memblock.memory, nid);
835                 node_set_online(nid);
836         }
837 }
838 
839 void __init dump_numa_cpu_topology(void)
840 {
841         unsigned int node;
842         unsigned int cpu, count;
843 
844         if (min_common_depth == -1 || !numa_enabled)
845                 return;
846 
847         for_each_online_node(node) {
848                 printk(KERN_DEBUG "Node %d CPUs:", node);
849 
850                 count = 0;
851                 /*
852                  * If we used a CPU iterator here we would miss printing
853                  * the holes in the cpumap.
854                  */
855                 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
856                         if (cpumask_test_cpu(cpu,
857                                         node_to_cpumask_map[node])) {
858                                 if (count == 0)
859                                         printk(" %u", cpu);
860                                 ++count;
861                         } else {
862                                 if (count > 1)
863                                         printk("-%u", cpu - 1);
864                                 count = 0;
865                         }
866                 }
867 
868                 if (count > 1)
869                         printk("-%u", nr_cpu_ids - 1);
870                 printk("\n");
871         }
872 }
873 
874 static void __init dump_numa_memory_topology(void)
875 {
876         unsigned int node;
877         unsigned int count;
878 
879         if (min_common_depth == -1 || !numa_enabled)
880                 return;
881 
882         for_each_online_node(node) {
883                 unsigned long i;
884 
885                 printk(KERN_DEBUG "Node %d Memory:", node);
886 
887                 count = 0;
888 
889                 for (i = 0; i < memblock_end_of_DRAM();
890                      i += (1 << SECTION_SIZE_BITS)) {
891                         if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
892                                 if (count == 0)
893                                         printk(" 0x%lx", i);
894                                 ++count;
895                         } else {
896                                 if (count > 0)
897                                         printk("-0x%lx", i);
898                                 count = 0;
899                         }
900                 }
901 
902                 if (count > 0)
903                         printk("-0x%lx", i);
904                 printk("\n");
905         }
906 }
907 
908 /* Initialize NODE_DATA for a node on the local memory */
909 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
910 {
911         u64 spanned_pages = end_pfn - start_pfn;
912         const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
913         u64 nd_pa;
914         void *nd;
915         int tnid;
916 
917         if (spanned_pages)
918                 pr_info("Initmem setup node %d [mem %#010Lx-%#010Lx]\n",
919                         nid, start_pfn << PAGE_SHIFT,
920                         (end_pfn << PAGE_SHIFT) - 1);
921         else
922                 pr_info("Initmem setup node %d\n", nid);
923 
924         nd_pa = memblock_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
925         nd = __va(nd_pa);
926 
927         /* report and initialize */
928         pr_info("  NODE_DATA [mem %#010Lx-%#010Lx]\n",
929                 nd_pa, nd_pa + nd_size - 1);
930         tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
931         if (tnid != nid)
932                 pr_info("    NODE_DATA(%d) on node %d\n", nid, tnid);
933 
934         node_data[nid] = nd;
935         memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
936         NODE_DATA(nid)->node_id = nid;
937         NODE_DATA(nid)->node_start_pfn = start_pfn;
938         NODE_DATA(nid)->node_spanned_pages = spanned_pages;
939 }
940 
941 void __init initmem_init(void)
942 {
943         int nid, cpu;
944 
945         max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
946         max_pfn = max_low_pfn;
947 
948         if (parse_numa_properties())
949                 setup_nonnuma();
950         else
951                 dump_numa_memory_topology();
952 
953         memblock_dump_all();
954 
955         /*
956          * Reduce the possible NUMA nodes to the online NUMA nodes,
957          * since we do not support node hotplug. This ensures that  we
958          * lower the maximum NUMA node ID to what is actually present.
959          */
960         nodes_and(node_possible_map, node_possible_map, node_online_map);
961 
962         for_each_online_node(nid) {
963                 unsigned long start_pfn, end_pfn;
964 
965                 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
966                 setup_node_data(nid, start_pfn, end_pfn);
967                 sparse_memory_present_with_active_regions(nid);
968         }
969 
970         sparse_init();
971 
972         setup_node_to_cpumask_map();
973 
974         reset_numa_cpu_lookup_table();
975 
976         /*
977          * We need the numa_cpu_lookup_table to be accurate for all CPUs,
978          * even before we online them, so that we can use cpu_to_{node,mem}
979          * early in boot, cf. smp_prepare_cpus().
980          * _nocalls() + manual invocation is used because cpuhp is not yet
981          * initialized for the boot CPU.
982          */
983         cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "POWER_NUMA_PREPARE",
984                                   ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
985         for_each_present_cpu(cpu)
986                 numa_setup_cpu(cpu);
987 }
988 
989 static int __init early_numa(char *p)
990 {
991         if (!p)
992                 return 0;
993 
994         if (strstr(p, "off"))
995                 numa_enabled = 0;
996 
997         if (strstr(p, "debug"))
998                 numa_debug = 1;
999 
1000         p = strstr(p, "fake=");
1001         if (p)
1002                 cmdline = p + strlen("fake=");
1003 
1004         return 0;
1005 }
1006 early_param("numa", early_numa);
1007 
1008 static bool topology_updates_enabled = true;
1009 
1010 static int __init early_topology_updates(char *p)
1011 {
1012         if (!p)
1013                 return 0;
1014 
1015         if (!strcmp(p, "off")) {
1016                 pr_info("Disabling topology updates\n");
1017                 topology_updates_enabled = false;
1018         }
1019 
1020         return 0;
1021 }
1022 early_param("topology_updates", early_topology_updates);
1023 
1024 #ifdef CONFIG_MEMORY_HOTPLUG
1025 /*
1026  * Find the node associated with a hot added memory section for
1027  * memory represented in the device tree by the property
1028  * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1029  */
1030 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1031                                      unsigned long scn_addr)
1032 {
1033         const __be32 *dm;
1034         unsigned int drconf_cell_cnt, rc;
1035         unsigned long lmb_size;
1036         struct assoc_arrays aa;
1037         int nid = -1;
1038 
1039         drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1040         if (!drconf_cell_cnt)
1041                 return -1;
1042 
1043         lmb_size = of_get_lmb_size(memory);
1044         if (!lmb_size)
1045                 return -1;
1046 
1047         rc = of_get_assoc_arrays(memory, &aa);
1048         if (rc)
1049                 return -1;
1050 
1051         for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1052                 struct of_drconf_cell drmem;
1053 
1054                 read_drconf_cell(&drmem, &dm);
1055 
1056                 /* skip this block if it is reserved or not assigned to
1057                  * this partition */
1058                 if ((drmem.flags & DRCONF_MEM_RESERVED)
1059                     || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1060                         continue;
1061 
1062                 if ((scn_addr < drmem.base_addr)
1063                     || (scn_addr >= (drmem.base_addr + lmb_size)))
1064                         continue;
1065 
1066                 nid = of_drconf_to_nid_single(&drmem, &aa);
1067                 break;
1068         }
1069 
1070         return nid;
1071 }
1072 
1073 /*
1074  * Find the node associated with a hot added memory section for memory
1075  * represented in the device tree as a node (i.e. memory@XXXX) for
1076  * each memblock.
1077  */
1078 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1079 {
1080         struct device_node *memory;
1081         int nid = -1;
1082 
1083         for_each_node_by_type(memory, "memory") {
1084                 unsigned long start, size;
1085                 int ranges;
1086                 const __be32 *memcell_buf;
1087                 unsigned int len;
1088 
1089                 memcell_buf = of_get_property(memory, "reg", &len);
1090                 if (!memcell_buf || len <= 0)
1091                         continue;
1092 
1093                 /* ranges in cell */
1094                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1095 
1096                 while (ranges--) {
1097                         start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1098                         size = read_n_cells(n_mem_size_cells, &memcell_buf);
1099 
1100                         if ((scn_addr < start) || (scn_addr >= (start + size)))
1101                                 continue;
1102 
1103                         nid = of_node_to_nid_single(memory);
1104                         break;
1105                 }
1106 
1107                 if (nid >= 0)
1108                         break;
1109         }
1110 
1111         of_node_put(memory);
1112 
1113         return nid;
1114 }
1115 
1116 /*
1117  * Find the node associated with a hot added memory section.  Section
1118  * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1119  * sections are fully contained within a single MEMBLOCK.
1120  */
1121 int hot_add_scn_to_nid(unsigned long scn_addr)
1122 {
1123         struct device_node *memory = NULL;
1124         int nid, found = 0;
1125 
1126         if (!numa_enabled || (min_common_depth < 0))
1127                 return first_online_node;
1128 
1129         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1130         if (memory) {
1131                 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1132                 of_node_put(memory);
1133         } else {
1134                 nid = hot_add_node_scn_to_nid(scn_addr);
1135         }
1136 
1137         if (nid < 0 || !node_online(nid))
1138                 nid = first_online_node;
1139 
1140         if (NODE_DATA(nid)->node_spanned_pages)
1141                 return nid;
1142 
1143         for_each_online_node(nid) {
1144                 if (NODE_DATA(nid)->node_spanned_pages) {
1145                         found = 1;
1146                         break;
1147                 }
1148         }
1149 
1150         BUG_ON(!found);
1151         return nid;
1152 }
1153 
1154 static u64 hot_add_drconf_memory_max(void)
1155 {
1156         struct device_node *memory = NULL;
1157         struct device_node *dn = NULL;
1158         unsigned int drconf_cell_cnt = 0;
1159         u64 lmb_size = 0;
1160         const __be32 *dm = NULL;
1161         const __be64 *lrdr = NULL;
1162         struct of_drconf_cell drmem;
1163 
1164         dn = of_find_node_by_path("/rtas");
1165         if (dn) {
1166                 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1167                 of_node_put(dn);
1168                 if (lrdr)
1169                         return be64_to_cpup(lrdr);
1170         }
1171 
1172         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1173         if (memory) {
1174                 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1175                 lmb_size = of_get_lmb_size(memory);
1176 
1177                 /* Advance to the last cell, each cell has 6 32 bit integers */
1178                 dm += (drconf_cell_cnt - 1) * 6;
1179                 read_drconf_cell(&drmem, &dm);
1180                 of_node_put(memory);
1181                 return drmem.base_addr + lmb_size;
1182         }
1183         return 0;
1184 }
1185 
1186 /*
1187  * memory_hotplug_max - return max address of memory that may be added
1188  *
1189  * This is currently only used on systems that support drconfig memory
1190  * hotplug.
1191  */
1192 u64 memory_hotplug_max(void)
1193 {
1194         return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1195 }
1196 #endif /* CONFIG_MEMORY_HOTPLUG */
1197 
1198 /* Virtual Processor Home Node (VPHN) support */
1199 #ifdef CONFIG_PPC_SPLPAR
1200 
1201 #include "vphn.h"
1202 
1203 struct topology_update_data {
1204         struct topology_update_data *next;
1205         unsigned int cpu;
1206         int old_nid;
1207         int new_nid;
1208 };
1209 
1210 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1211 static cpumask_t cpu_associativity_changes_mask;
1212 static int vphn_enabled;
1213 static int prrn_enabled;
1214 static void reset_topology_timer(void);
1215 
1216 /*
1217  * Store the current values of the associativity change counters in the
1218  * hypervisor.
1219  */
1220 static void setup_cpu_associativity_change_counters(void)
1221 {
1222         int cpu;
1223 
1224         /* The VPHN feature supports a maximum of 8 reference points */
1225         BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1226 
1227         for_each_possible_cpu(cpu) {
1228                 int i;
1229                 u8 *counts = vphn_cpu_change_counts[cpu];
1230                 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1231 
1232                 for (i = 0; i < distance_ref_points_depth; i++)
1233                         counts[i] = hypervisor_counts[i];
1234         }
1235 }
1236 
1237 /*
1238  * The hypervisor maintains a set of 8 associativity change counters in
1239  * the VPA of each cpu that correspond to the associativity levels in the
1240  * ibm,associativity-reference-points property. When an associativity
1241  * level changes, the corresponding counter is incremented.
1242  *
1243  * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1244  * node associativity levels have changed.
1245  *
1246  * Returns the number of cpus with unhandled associativity changes.
1247  */
1248 static int update_cpu_associativity_changes_mask(void)
1249 {
1250         int cpu;
1251         cpumask_t *changes = &cpu_associativity_changes_mask;
1252 
1253         for_each_possible_cpu(cpu) {
1254                 int i, changed = 0;
1255                 u8 *counts = vphn_cpu_change_counts[cpu];
1256                 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1257 
1258                 for (i = 0; i < distance_ref_points_depth; i++) {
1259                         if (hypervisor_counts[i] != counts[i]) {
1260                                 counts[i] = hypervisor_counts[i];
1261                                 changed = 1;
1262                         }
1263                 }
1264                 if (changed) {
1265                         cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1266                         cpu = cpu_last_thread_sibling(cpu);
1267                 }
1268         }
1269 
1270         return cpumask_weight(changes);
1271 }
1272 
1273 /*
1274  * Retrieve the new associativity information for a virtual processor's
1275  * home node.
1276  */
1277 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1278 {
1279         long rc;
1280         long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1281         u64 flags = 1;
1282         int hwcpu = get_hard_smp_processor_id(cpu);
1283 
1284         rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1285         vphn_unpack_associativity(retbuf, associativity);
1286 
1287         return rc;
1288 }
1289 
1290 static long vphn_get_associativity(unsigned long cpu,
1291                                         __be32 *associativity)
1292 {
1293         long rc;
1294 
1295         rc = hcall_vphn(cpu, associativity);
1296 
1297         switch (rc) {
1298         case H_FUNCTION:
1299                 printk(KERN_INFO
1300                         "VPHN is not supported. Disabling polling...\n");
1301                 stop_topology_update();
1302                 break;
1303         case H_HARDWARE:
1304                 printk(KERN_ERR
1305                         "hcall_vphn() experienced a hardware fault "
1306                         "preventing VPHN. Disabling polling...\n");
1307                 stop_topology_update();
1308         }
1309 
1310         return rc;
1311 }
1312 
1313 /*
1314  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1315  * characteristics change. This function doesn't perform any locking and is
1316  * only safe to call from stop_machine().
1317  */
1318 static int update_cpu_topology(void *data)
1319 {
1320         struct topology_update_data *update;
1321         unsigned long cpu;
1322 
1323         if (!data)
1324                 return -EINVAL;
1325 
1326         cpu = smp_processor_id();
1327 
1328         for (update = data; update; update = update->next) {
1329                 int new_nid = update->new_nid;
1330                 if (cpu != update->cpu)
1331                         continue;
1332 
1333                 unmap_cpu_from_node(cpu);
1334                 map_cpu_to_node(cpu, new_nid);
1335                 set_cpu_numa_node(cpu, new_nid);
1336                 set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1337                 vdso_getcpu_init();
1338         }
1339 
1340         return 0;
1341 }
1342 
1343 static int update_lookup_table(void *data)
1344 {
1345         struct topology_update_data *update;
1346 
1347         if (!data)
1348                 return -EINVAL;
1349 
1350         /*
1351          * Upon topology update, the numa-cpu lookup table needs to be updated
1352          * for all threads in the core, including offline CPUs, to ensure that
1353          * future hotplug operations respect the cpu-to-node associativity
1354          * properly.
1355          */
1356         for (update = data; update; update = update->next) {
1357                 int nid, base, j;
1358 
1359                 nid = update->new_nid;
1360                 base = cpu_first_thread_sibling(update->cpu);
1361 
1362                 for (j = 0; j < threads_per_core; j++) {
1363                         update_numa_cpu_lookup_table(base + j, nid);
1364                 }
1365         }
1366 
1367         return 0;
1368 }
1369 
1370 /*
1371  * Update the node maps and sysfs entries for each cpu whose home node
1372  * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1373  */
1374 int arch_update_cpu_topology(void)
1375 {
1376         unsigned int cpu, sibling, changed = 0;
1377         struct topology_update_data *updates, *ud;
1378         __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1379         cpumask_t updated_cpus;
1380         struct device *dev;
1381         int weight, new_nid, i = 0;
1382 
1383         if (!prrn_enabled && !vphn_enabled)
1384                 return 0;
1385 
1386         weight = cpumask_weight(&cpu_associativity_changes_mask);
1387         if (!weight)
1388                 return 0;
1389 
1390         updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1391         if (!updates)
1392                 return 0;
1393 
1394         cpumask_clear(&updated_cpus);
1395 
1396         for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1397                 /*
1398                  * If siblings aren't flagged for changes, updates list
1399                  * will be too short. Skip on this update and set for next
1400                  * update.
1401                  */
1402                 if (!cpumask_subset(cpu_sibling_mask(cpu),
1403                                         &cpu_associativity_changes_mask)) {
1404                         pr_info("Sibling bits not set for associativity "
1405                                         "change, cpu%d\n", cpu);
1406                         cpumask_or(&cpu_associativity_changes_mask,
1407                                         &cpu_associativity_changes_mask,
1408                                         cpu_sibling_mask(cpu));
1409                         cpu = cpu_last_thread_sibling(cpu);
1410                         continue;
1411                 }
1412 
1413                 /* Use associativity from first thread for all siblings */
1414                 vphn_get_associativity(cpu, associativity);
1415                 new_nid = associativity_to_nid(associativity);
1416                 if (new_nid < 0 || !node_online(new_nid))
1417                         new_nid = first_online_node;
1418 
1419                 if (new_nid == numa_cpu_lookup_table[cpu]) {
1420                         cpumask_andnot(&cpu_associativity_changes_mask,
1421                                         &cpu_associativity_changes_mask,
1422                                         cpu_sibling_mask(cpu));
1423                         cpu = cpu_last_thread_sibling(cpu);
1424                         continue;
1425                 }
1426 
1427                 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1428                         ud = &updates[i++];
1429                         ud->cpu = sibling;
1430                         ud->new_nid = new_nid;
1431                         ud->old_nid = numa_cpu_lookup_table[sibling];
1432                         cpumask_set_cpu(sibling, &updated_cpus);
1433                         if (i < weight)
1434                                 ud->next = &updates[i];
1435                 }
1436                 cpu = cpu_last_thread_sibling(cpu);
1437         }
1438 
1439         pr_debug("Topology update for the following CPUs:\n");
1440         if (cpumask_weight(&updated_cpus)) {
1441                 for (ud = &updates[0]; ud; ud = ud->next) {
1442                         pr_debug("cpu %d moving from node %d "
1443                                           "to %d\n", ud->cpu,
1444                                           ud->old_nid, ud->new_nid);
1445                 }
1446         }
1447 
1448         /*
1449          * In cases where we have nothing to update (because the updates list
1450          * is too short or because the new topology is same as the old one),
1451          * skip invoking update_cpu_topology() via stop-machine(). This is
1452          * necessary (and not just a fast-path optimization) since stop-machine
1453          * can end up electing a random CPU to run update_cpu_topology(), and
1454          * thus trick us into setting up incorrect cpu-node mappings (since
1455          * 'updates' is kzalloc()'ed).
1456          *
1457          * And for the similar reason, we will skip all the following updating.
1458          */
1459         if (!cpumask_weight(&updated_cpus))
1460                 goto out;
1461 
1462         stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1463 
1464         /*
1465          * Update the numa-cpu lookup table with the new mappings, even for
1466          * offline CPUs. It is best to perform this update from the stop-
1467          * machine context.
1468          */
1469         stop_machine(update_lookup_table, &updates[0],
1470                                         cpumask_of(raw_smp_processor_id()));
1471 
1472         for (ud = &updates[0]; ud; ud = ud->next) {
1473                 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1474                 register_cpu_under_node(ud->cpu, ud->new_nid);
1475 
1476                 dev = get_cpu_device(ud->cpu);
1477                 if (dev)
1478                         kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1479                 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1480                 changed = 1;
1481         }
1482 
1483 out:
1484         kfree(updates);
1485         return changed;
1486 }
1487 
1488 static void topology_work_fn(struct work_struct *work)
1489 {
1490         rebuild_sched_domains();
1491 }
1492 static DECLARE_WORK(topology_work, topology_work_fn);
1493 
1494 static void topology_schedule_update(void)
1495 {
1496         schedule_work(&topology_work);
1497 }
1498 
1499 static void topology_timer_fn(unsigned long ignored)
1500 {
1501         if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1502                 topology_schedule_update();
1503         else if (vphn_enabled) {
1504                 if (update_cpu_associativity_changes_mask() > 0)
1505                         topology_schedule_update();
1506                 reset_topology_timer();
1507         }
1508 }
1509 static struct timer_list topology_timer =
1510         TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1511 
1512 static void reset_topology_timer(void)
1513 {
1514         topology_timer.data = 0;
1515         topology_timer.expires = jiffies + 60 * HZ;
1516         mod_timer(&topology_timer, topology_timer.expires);
1517 }
1518 
1519 #ifdef CONFIG_SMP
1520 
1521 static void stage_topology_update(int core_id)
1522 {
1523         cpumask_or(&cpu_associativity_changes_mask,
1524                 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1525         reset_topology_timer();
1526 }
1527 
1528 static int dt_update_callback(struct notifier_block *nb,
1529                                 unsigned long action, void *data)
1530 {
1531         struct of_reconfig_data *update = data;
1532         int rc = NOTIFY_DONE;
1533 
1534         switch (action) {
1535         case OF_RECONFIG_UPDATE_PROPERTY:
1536                 if (!of_prop_cmp(update->dn->type, "cpu") &&
1537                     !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1538                         u32 core_id;
1539                         of_property_read_u32(update->dn, "reg", &core_id);
1540                         stage_topology_update(core_id);
1541                         rc = NOTIFY_OK;
1542                 }
1543                 break;
1544         }
1545 
1546         return rc;
1547 }
1548 
1549 static struct notifier_block dt_update_nb = {
1550         .notifier_call = dt_update_callback,
1551 };
1552 
1553 #endif
1554 
1555 /*
1556  * Start polling for associativity changes.
1557  */
1558 int start_topology_update(void)
1559 {
1560         int rc = 0;
1561 
1562         if (firmware_has_feature(FW_FEATURE_PRRN)) {
1563                 if (!prrn_enabled) {
1564                         prrn_enabled = 1;
1565                         vphn_enabled = 0;
1566 #ifdef CONFIG_SMP
1567                         rc = of_reconfig_notifier_register(&dt_update_nb);
1568 #endif
1569                 }
1570         } else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1571                    lppaca_shared_proc(get_lppaca())) {
1572                 if (!vphn_enabled) {
1573                         prrn_enabled = 0;
1574                         vphn_enabled = 1;
1575                         setup_cpu_associativity_change_counters();
1576                         init_timer_deferrable(&topology_timer);
1577                         reset_topology_timer();
1578                 }
1579         }
1580 
1581         return rc;
1582 }
1583 
1584 /*
1585  * Disable polling for VPHN associativity changes.
1586  */
1587 int stop_topology_update(void)
1588 {
1589         int rc = 0;
1590 
1591         if (prrn_enabled) {
1592                 prrn_enabled = 0;
1593 #ifdef CONFIG_SMP
1594                 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1595 #endif
1596         } else if (vphn_enabled) {
1597                 vphn_enabled = 0;
1598                 rc = del_timer_sync(&topology_timer);
1599         }
1600 
1601         return rc;
1602 }
1603 
1604 int prrn_is_enabled(void)
1605 {
1606         return prrn_enabled;
1607 }
1608 
1609 static int topology_read(struct seq_file *file, void *v)
1610 {
1611         if (vphn_enabled || prrn_enabled)
1612                 seq_puts(file, "on\n");
1613         else
1614                 seq_puts(file, "off\n");
1615 
1616         return 0;
1617 }
1618 
1619 static int topology_open(struct inode *inode, struct file *file)
1620 {
1621         return single_open(file, topology_read, NULL);
1622 }
1623 
1624 static ssize_t topology_write(struct file *file, const char __user *buf,
1625                               size_t count, loff_t *off)
1626 {
1627         char kbuf[4]; /* "on" or "off" plus null. */
1628         int read_len;
1629 
1630         read_len = count < 3 ? count : 3;
1631         if (copy_from_user(kbuf, buf, read_len))
1632                 return -EINVAL;
1633 
1634         kbuf[read_len] = '\0';
1635 
1636         if (!strncmp(kbuf, "on", 2))
1637                 start_topology_update();
1638         else if (!strncmp(kbuf, "off", 3))
1639                 stop_topology_update();
1640         else
1641                 return -EINVAL;
1642 
1643         return count;
1644 }
1645 
1646 static const struct file_operations topology_ops = {
1647         .read = seq_read,
1648         .write = topology_write,
1649         .open = topology_open,
1650         .release = single_release
1651 };
1652 
1653 static int topology_update_init(void)
1654 {
1655         /* Do not poll for changes if disabled at boot */
1656         if (topology_updates_enabled)
1657                 start_topology_update();
1658 
1659         if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1660                 return -ENOMEM;
1661 
1662         return 0;
1663 }
1664 device_initcall(topology_update_init);
1665 #endif /* CONFIG_PPC_SPLPAR */
1666 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp