| 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 | static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action, |
| 585 | void *hcpu) |
| 586 | { |
| 587 | unsigned long lcpu = (unsigned long)hcpu; |
| 588 | int ret = NOTIFY_DONE, nid; |
| 589 | |
| 590 | switch (action) { |
| 591 | case CPU_UP_PREPARE: |
| 592 | case CPU_UP_PREPARE_FROZEN: |
| 593 | nid = numa_setup_cpu(lcpu); |
| 594 | verify_cpu_node_mapping((int)lcpu, nid); |
| 595 | ret = NOTIFY_OK; |
| 596 | break; |
| 597 | #ifdef CONFIG_HOTPLUG_CPU |
| 598 | case CPU_DEAD: |
| 599 | case CPU_DEAD_FROZEN: |
| 600 | case CPU_UP_CANCELED: |
| 601 | case CPU_UP_CANCELED_FROZEN: |
| 602 | unmap_cpu_from_node(lcpu); |
| 603 | ret = NOTIFY_OK; |
| 604 | break; |
| 605 | #endif |
| 606 | } |
| 607 | return ret; |
| 608 | } |
| 609 | |
| 610 | /* |
| 611 | * Check and possibly modify a memory region to enforce the memory limit. |
| 612 | * |
| 613 | * Returns the size the region should have to enforce the memory limit. |
| 614 | * This will either be the original value of size, a truncated value, |
| 615 | * or zero. If the returned value of size is 0 the region should be |
| 616 | * discarded as it lies wholly above the memory limit. |
| 617 | */ |
| 618 | static unsigned long __init numa_enforce_memory_limit(unsigned long start, |
| 619 | unsigned long size) |
| 620 | { |
| 621 | /* |
| 622 | * We use memblock_end_of_DRAM() in here instead of memory_limit because |
| 623 | * we've already adjusted it for the limit and it takes care of |
| 624 | * having memory holes below the limit. Also, in the case of |
| 625 | * iommu_is_off, memory_limit is not set but is implicitly enforced. |
| 626 | */ |
| 627 | |
| 628 | if (start + size <= memblock_end_of_DRAM()) |
| 629 | return size; |
| 630 | |
| 631 | if (start >= memblock_end_of_DRAM()) |
| 632 | return 0; |
| 633 | |
| 634 | return memblock_end_of_DRAM() - start; |
| 635 | } |
| 636 | |
| 637 | /* |
| 638 | * Reads the counter for a given entry in |
| 639 | * linux,drconf-usable-memory property |
| 640 | */ |
| 641 | static inline int __init read_usm_ranges(const __be32 **usm) |
| 642 | { |
| 643 | /* |
| 644 | * For each lmb in ibm,dynamic-memory a corresponding |
| 645 | * entry in linux,drconf-usable-memory property contains |
| 646 | * a counter followed by that many (base, size) duple. |
| 647 | * read the counter from linux,drconf-usable-memory |
| 648 | */ |
| 649 | return read_n_cells(n_mem_size_cells, usm); |
| 650 | } |
| 651 | |
| 652 | /* |
| 653 | * Extract NUMA information from the ibm,dynamic-reconfiguration-memory |
| 654 | * node. This assumes n_mem_{addr,size}_cells have been set. |
| 655 | */ |
| 656 | static void __init parse_drconf_memory(struct device_node *memory) |
| 657 | { |
| 658 | const __be32 *uninitialized_var(dm), *usm; |
| 659 | unsigned int n, rc, ranges, is_kexec_kdump = 0; |
| 660 | unsigned long lmb_size, base, size, sz; |
| 661 | int nid; |
| 662 | struct assoc_arrays aa = { .arrays = NULL }; |
| 663 | |
| 664 | n = of_get_drconf_memory(memory, &dm); |
| 665 | if (!n) |
| 666 | return; |
| 667 | |
| 668 | lmb_size = of_get_lmb_size(memory); |
| 669 | if (!lmb_size) |
| 670 | return; |
| 671 | |
| 672 | rc = of_get_assoc_arrays(memory, &aa); |
| 673 | if (rc) |
| 674 | return; |
| 675 | |
| 676 | /* check if this is a kexec/kdump kernel */ |
| 677 | usm = of_get_usable_memory(memory); |
| 678 | if (usm != NULL) |
| 679 | is_kexec_kdump = 1; |
| 680 | |
| 681 | for (; n != 0; --n) { |
| 682 | struct of_drconf_cell drmem; |
| 683 | |
| 684 | read_drconf_cell(&drmem, &dm); |
| 685 | |
| 686 | /* skip this block if the reserved bit is set in flags (0x80) |
| 687 | or if the block is not assigned to this partition (0x8) */ |
| 688 | if ((drmem.flags & DRCONF_MEM_RESERVED) |
| 689 | || !(drmem.flags & DRCONF_MEM_ASSIGNED)) |
| 690 | continue; |
| 691 | |
| 692 | base = drmem.base_addr; |
| 693 | size = lmb_size; |
| 694 | ranges = 1; |
| 695 | |
| 696 | if (is_kexec_kdump) { |
| 697 | ranges = read_usm_ranges(&usm); |
| 698 | if (!ranges) /* there are no (base, size) duple */ |
| 699 | continue; |
| 700 | } |
| 701 | do { |
| 702 | if (is_kexec_kdump) { |
| 703 | base = read_n_cells(n_mem_addr_cells, &usm); |
| 704 | size = read_n_cells(n_mem_size_cells, &usm); |
| 705 | } |
| 706 | nid = of_drconf_to_nid_single(&drmem, &aa); |
| 707 | fake_numa_create_new_node( |
| 708 | ((base + size) >> PAGE_SHIFT), |
| 709 | &nid); |
| 710 | node_set_online(nid); |
| 711 | sz = numa_enforce_memory_limit(base, size); |
| 712 | if (sz) |
| 713 | memblock_set_node(base, sz, |
| 714 | &memblock.memory, nid); |
| 715 | } while (--ranges); |
| 716 | } |
| 717 | } |
| 718 | |
| 719 | static int __init parse_numa_properties(void) |
| 720 | { |
| 721 | struct device_node *memory; |
| 722 | int default_nid = 0; |
| 723 | unsigned long i; |
| 724 | |
| 725 | if (numa_enabled == 0) { |
| 726 | printk(KERN_WARNING "NUMA disabled by user\n"); |
| 727 | return -1; |
| 728 | } |
| 729 | |
| 730 | min_common_depth = find_min_common_depth(); |
| 731 | |
| 732 | if (min_common_depth < 0) |
| 733 | return min_common_depth; |
| 734 | |
| 735 | dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth); |
| 736 | |
| 737 | /* |
| 738 | * Even though we connect cpus to numa domains later in SMP |
| 739 | * init, we need to know the node ids now. This is because |
| 740 | * each node to be onlined must have NODE_DATA etc backing it. |
| 741 | */ |
| 742 | for_each_present_cpu(i) { |
| 743 | struct device_node *cpu; |
| 744 | int nid; |
| 745 | |
| 746 | cpu = of_get_cpu_node(i, NULL); |
| 747 | BUG_ON(!cpu); |
| 748 | nid = of_node_to_nid_single(cpu); |
| 749 | of_node_put(cpu); |
| 750 | |
| 751 | /* |
| 752 | * Don't fall back to default_nid yet -- we will plug |
| 753 | * cpus into nodes once the memory scan has discovered |
| 754 | * the topology. |
| 755 | */ |
| 756 | if (nid < 0) |
| 757 | continue; |
| 758 | node_set_online(nid); |
| 759 | } |
| 760 | |
| 761 | get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells); |
| 762 | |
| 763 | for_each_node_by_type(memory, "memory") { |
| 764 | unsigned long start; |
| 765 | unsigned long size; |
| 766 | int nid; |
| 767 | int ranges; |
| 768 | const __be32 *memcell_buf; |
| 769 | unsigned int len; |
| 770 | |
| 771 | memcell_buf = of_get_property(memory, |
| 772 | "linux,usable-memory", &len); |
| 773 | if (!memcell_buf || len <= 0) |
| 774 | memcell_buf = of_get_property(memory, "reg", &len); |
| 775 | if (!memcell_buf || len <= 0) |
| 776 | continue; |
| 777 | |
| 778 | /* ranges in cell */ |
| 779 | ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells); |
| 780 | new_range: |
| 781 | /* these are order-sensitive, and modify the buffer pointer */ |
| 782 | start = read_n_cells(n_mem_addr_cells, &memcell_buf); |
| 783 | size = read_n_cells(n_mem_size_cells, &memcell_buf); |
| 784 | |
| 785 | /* |
| 786 | * Assumption: either all memory nodes or none will |
| 787 | * have associativity properties. If none, then |
| 788 | * everything goes to default_nid. |
| 789 | */ |
| 790 | nid = of_node_to_nid_single(memory); |
| 791 | if (nid < 0) |
| 792 | nid = default_nid; |
| 793 | |
| 794 | fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid); |
| 795 | node_set_online(nid); |
| 796 | |
| 797 | if (!(size = numa_enforce_memory_limit(start, size))) { |
| 798 | if (--ranges) |
| 799 | goto new_range; |
| 800 | else |
| 801 | continue; |
| 802 | } |
| 803 | |
| 804 | memblock_set_node(start, size, &memblock.memory, nid); |
| 805 | |
| 806 | if (--ranges) |
| 807 | goto new_range; |
| 808 | } |
| 809 | |
| 810 | /* |
| 811 | * Now do the same thing for each MEMBLOCK listed in the |
| 812 | * ibm,dynamic-memory property in the |
| 813 | * ibm,dynamic-reconfiguration-memory node. |
| 814 | */ |
| 815 | memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); |
| 816 | if (memory) |
| 817 | parse_drconf_memory(memory); |
| 818 | |
| 819 | return 0; |
| 820 | } |
| 821 | |
| 822 | static void __init setup_nonnuma(void) |
| 823 | { |
| 824 | unsigned long top_of_ram = memblock_end_of_DRAM(); |
| 825 | unsigned long total_ram = memblock_phys_mem_size(); |
| 826 | unsigned long start_pfn, end_pfn; |
| 827 | unsigned int nid = 0; |
| 828 | struct memblock_region *reg; |
| 829 | |
| 830 | printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n", |
| 831 | top_of_ram, total_ram); |
| 832 | printk(KERN_DEBUG "Memory hole size: %ldMB\n", |
| 833 | (top_of_ram - total_ram) >> 20); |
| 834 | |
| 835 | for_each_memblock(memory, reg) { |
| 836 | start_pfn = memblock_region_memory_base_pfn(reg); |
| 837 | end_pfn = memblock_region_memory_end_pfn(reg); |
| 838 | |
| 839 | fake_numa_create_new_node(end_pfn, &nid); |
| 840 | memblock_set_node(PFN_PHYS(start_pfn), |
| 841 | PFN_PHYS(end_pfn - start_pfn), |
| 842 | &memblock.memory, nid); |
| 843 | node_set_online(nid); |
| 844 | } |
| 845 | } |
| 846 | |
| 847 | void __init dump_numa_cpu_topology(void) |
| 848 | { |
| 849 | unsigned int node; |
| 850 | unsigned int cpu, count; |
| 851 | |
| 852 | if (min_common_depth == -1 || !numa_enabled) |
| 853 | return; |
| 854 | |
| 855 | for_each_online_node(node) { |
| 856 | printk(KERN_DEBUG "Node %d CPUs:", node); |
| 857 | |
| 858 | count = 0; |
| 859 | /* |
| 860 | * If we used a CPU iterator here we would miss printing |
| 861 | * the holes in the cpumap. |
| 862 | */ |
| 863 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) { |
| 864 | if (cpumask_test_cpu(cpu, |
| 865 | node_to_cpumask_map[node])) { |
| 866 | if (count == 0) |
| 867 | printk(" %u", cpu); |
| 868 | ++count; |
| 869 | } else { |
| 870 | if (count > 1) |
| 871 | printk("-%u", cpu - 1); |
| 872 | count = 0; |
| 873 | } |
| 874 | } |
| 875 | |
| 876 | if (count > 1) |
| 877 | printk("-%u", nr_cpu_ids - 1); |
| 878 | printk("\n"); |
| 879 | } |
| 880 | } |
| 881 | |
| 882 | static void __init dump_numa_memory_topology(void) |
| 883 | { |
| 884 | unsigned int node; |
| 885 | unsigned int count; |
| 886 | |
| 887 | if (min_common_depth == -1 || !numa_enabled) |
| 888 | return; |
| 889 | |
| 890 | for_each_online_node(node) { |
| 891 | unsigned long i; |
| 892 | |
| 893 | printk(KERN_DEBUG "Node %d Memory:", node); |
| 894 | |
| 895 | count = 0; |
| 896 | |
| 897 | for (i = 0; i < memblock_end_of_DRAM(); |
| 898 | i += (1 << SECTION_SIZE_BITS)) { |
| 899 | if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) { |
| 900 | if (count == 0) |
| 901 | printk(" 0x%lx", i); |
| 902 | ++count; |
| 903 | } else { |
| 904 | if (count > 0) |
| 905 | printk("-0x%lx", i); |
| 906 | count = 0; |
| 907 | } |
| 908 | } |
| 909 | |
| 910 | if (count > 0) |
| 911 | printk("-0x%lx", i); |
| 912 | printk("\n"); |
| 913 | } |
| 914 | } |
| 915 | |
| 916 | static struct notifier_block ppc64_numa_nb = { |
| 917 | .notifier_call = cpu_numa_callback, |
| 918 | .priority = 1 /* Must run before sched domains notifier. */ |
| 919 | }; |
| 920 | |
| 921 | /* Initialize NODE_DATA for a node on the local memory */ |
| 922 | static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn) |
| 923 | { |
| 924 | u64 spanned_pages = end_pfn - start_pfn; |
| 925 | const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES); |
| 926 | u64 nd_pa; |
| 927 | void *nd; |
| 928 | int tnid; |
| 929 | |
| 930 | if (spanned_pages) |
| 931 | pr_info("Initmem setup node %d [mem %#010Lx-%#010Lx]\n", |
| 932 | nid, start_pfn << PAGE_SHIFT, |
| 933 | (end_pfn << PAGE_SHIFT) - 1); |
| 934 | else |
| 935 | pr_info("Initmem setup node %d\n", nid); |
| 936 | |
| 937 | nd_pa = memblock_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid); |
| 938 | nd = __va(nd_pa); |
| 939 | |
| 940 | /* report and initialize */ |
| 941 | pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n", |
| 942 | nd_pa, nd_pa + nd_size - 1); |
| 943 | tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT); |
| 944 | if (tnid != nid) |
| 945 | pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid); |
| 946 | |
| 947 | node_data[nid] = nd; |
| 948 | memset(NODE_DATA(nid), 0, sizeof(pg_data_t)); |
| 949 | NODE_DATA(nid)->node_id = nid; |
| 950 | NODE_DATA(nid)->node_start_pfn = start_pfn; |
| 951 | NODE_DATA(nid)->node_spanned_pages = spanned_pages; |
| 952 | } |
| 953 | |
| 954 | void __init initmem_init(void) |
| 955 | { |
| 956 | int nid, cpu; |
| 957 | |
| 958 | max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT; |
| 959 | max_pfn = max_low_pfn; |
| 960 | |
| 961 | if (parse_numa_properties()) |
| 962 | setup_nonnuma(); |
| 963 | else |
| 964 | dump_numa_memory_topology(); |
| 965 | |
| 966 | memblock_dump_all(); |
| 967 | |
| 968 | /* |
| 969 | * Reduce the possible NUMA nodes to the online NUMA nodes, |
| 970 | * since we do not support node hotplug. This ensures that we |
| 971 | * lower the maximum NUMA node ID to what is actually present. |
| 972 | */ |
| 973 | nodes_and(node_possible_map, node_possible_map, node_online_map); |
| 974 | |
| 975 | for_each_online_node(nid) { |
| 976 | unsigned long start_pfn, end_pfn; |
| 977 | |
| 978 | get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); |
| 979 | setup_node_data(nid, start_pfn, end_pfn); |
| 980 | sparse_memory_present_with_active_regions(nid); |
| 981 | } |
| 982 | |
| 983 | sparse_init(); |
| 984 | |
| 985 | setup_node_to_cpumask_map(); |
| 986 | |
| 987 | reset_numa_cpu_lookup_table(); |
| 988 | register_cpu_notifier(&ppc64_numa_nb); |
| 989 | /* |
| 990 | * We need the numa_cpu_lookup_table to be accurate for all CPUs, |
| 991 | * even before we online them, so that we can use cpu_to_{node,mem} |
| 992 | * early in boot, cf. smp_prepare_cpus(). |
| 993 | */ |
| 994 | for_each_present_cpu(cpu) { |
| 995 | numa_setup_cpu((unsigned long)cpu); |
| 996 | } |
| 997 | } |
| 998 | |
| 999 | static int __init early_numa(char *p) |
| 1000 | { |
| 1001 | if (!p) |
| 1002 | return 0; |
| 1003 | |
| 1004 | if (strstr(p, "off")) |
| 1005 | numa_enabled = 0; |
| 1006 | |
| 1007 | if (strstr(p, "debug")) |
| 1008 | numa_debug = 1; |
| 1009 | |
| 1010 | p = strstr(p, "fake="); |
| 1011 | if (p) |
| 1012 | cmdline = p + strlen("fake="); |
| 1013 | |
| 1014 | return 0; |
| 1015 | } |
| 1016 | early_param("numa", early_numa); |
| 1017 | |
| 1018 | static bool topology_updates_enabled = true; |
| 1019 | |
| 1020 | static int __init early_topology_updates(char *p) |
| 1021 | { |
| 1022 | if (!p) |
| 1023 | return 0; |
| 1024 | |
| 1025 | if (!strcmp(p, "off")) { |
| 1026 | pr_info("Disabling topology updates\n"); |
| 1027 | topology_updates_enabled = false; |
| 1028 | } |
| 1029 | |
| 1030 | return 0; |
| 1031 | } |
| 1032 | early_param("topology_updates", early_topology_updates); |
| 1033 | |
| 1034 | #ifdef CONFIG_MEMORY_HOTPLUG |
| 1035 | /* |
| 1036 | * Find the node associated with a hot added memory section for |
| 1037 | * memory represented in the device tree by the property |
| 1038 | * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory. |
| 1039 | */ |
| 1040 | static int hot_add_drconf_scn_to_nid(struct device_node *memory, |
| 1041 | unsigned long scn_addr) |
| 1042 | { |
| 1043 | const __be32 *dm; |
| 1044 | unsigned int drconf_cell_cnt, rc; |
| 1045 | unsigned long lmb_size; |
| 1046 | struct assoc_arrays aa; |
| 1047 | int nid = -1; |
| 1048 | |
| 1049 | drconf_cell_cnt = of_get_drconf_memory(memory, &dm); |
| 1050 | if (!drconf_cell_cnt) |
| 1051 | return -1; |
| 1052 | |
| 1053 | lmb_size = of_get_lmb_size(memory); |
| 1054 | if (!lmb_size) |
| 1055 | return -1; |
| 1056 | |
| 1057 | rc = of_get_assoc_arrays(memory, &aa); |
| 1058 | if (rc) |
| 1059 | return -1; |
| 1060 | |
| 1061 | for (; drconf_cell_cnt != 0; --drconf_cell_cnt) { |
| 1062 | struct of_drconf_cell drmem; |
| 1063 | |
| 1064 | read_drconf_cell(&drmem, &dm); |
| 1065 | |
| 1066 | /* skip this block if it is reserved or not assigned to |
| 1067 | * this partition */ |
| 1068 | if ((drmem.flags & DRCONF_MEM_RESERVED) |
| 1069 | || !(drmem.flags & DRCONF_MEM_ASSIGNED)) |
| 1070 | continue; |
| 1071 | |
| 1072 | if ((scn_addr < drmem.base_addr) |
| 1073 | || (scn_addr >= (drmem.base_addr + lmb_size))) |
| 1074 | continue; |
| 1075 | |
| 1076 | nid = of_drconf_to_nid_single(&drmem, &aa); |
| 1077 | break; |
| 1078 | } |
| 1079 | |
| 1080 | return nid; |
| 1081 | } |
| 1082 | |
| 1083 | /* |
| 1084 | * Find the node associated with a hot added memory section for memory |
| 1085 | * represented in the device tree as a node (i.e. memory@XXXX) for |
| 1086 | * each memblock. |
| 1087 | */ |
| 1088 | static int hot_add_node_scn_to_nid(unsigned long scn_addr) |
| 1089 | { |
| 1090 | struct device_node *memory; |
| 1091 | int nid = -1; |
| 1092 | |
| 1093 | for_each_node_by_type(memory, "memory") { |
| 1094 | unsigned long start, size; |
| 1095 | int ranges; |
| 1096 | const __be32 *memcell_buf; |
| 1097 | unsigned int len; |
| 1098 | |
| 1099 | memcell_buf = of_get_property(memory, "reg", &len); |
| 1100 | if (!memcell_buf || len <= 0) |
| 1101 | continue; |
| 1102 | |
| 1103 | /* ranges in cell */ |
| 1104 | ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells); |
| 1105 | |
| 1106 | while (ranges--) { |
| 1107 | start = read_n_cells(n_mem_addr_cells, &memcell_buf); |
| 1108 | size = read_n_cells(n_mem_size_cells, &memcell_buf); |
| 1109 | |
| 1110 | if ((scn_addr < start) || (scn_addr >= (start + size))) |
| 1111 | continue; |
| 1112 | |
| 1113 | nid = of_node_to_nid_single(memory); |
| 1114 | break; |
| 1115 | } |
| 1116 | |
| 1117 | if (nid >= 0) |
| 1118 | break; |
| 1119 | } |
| 1120 | |
| 1121 | of_node_put(memory); |
| 1122 | |
| 1123 | return nid; |
| 1124 | } |
| 1125 | |
| 1126 | /* |
| 1127 | * Find the node associated with a hot added memory section. Section |
| 1128 | * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that |
| 1129 | * sections are fully contained within a single MEMBLOCK. |
| 1130 | */ |
| 1131 | int hot_add_scn_to_nid(unsigned long scn_addr) |
| 1132 | { |
| 1133 | struct device_node *memory = NULL; |
| 1134 | int nid, found = 0; |
| 1135 | |
| 1136 | if (!numa_enabled || (min_common_depth < 0)) |
| 1137 | return first_online_node; |
| 1138 | |
| 1139 | memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); |
| 1140 | if (memory) { |
| 1141 | nid = hot_add_drconf_scn_to_nid(memory, scn_addr); |
| 1142 | of_node_put(memory); |
| 1143 | } else { |
| 1144 | nid = hot_add_node_scn_to_nid(scn_addr); |
| 1145 | } |
| 1146 | |
| 1147 | if (nid < 0 || !node_online(nid)) |
| 1148 | nid = first_online_node; |
| 1149 | |
| 1150 | if (NODE_DATA(nid)->node_spanned_pages) |
| 1151 | return nid; |
| 1152 | |
| 1153 | for_each_online_node(nid) { |
| 1154 | if (NODE_DATA(nid)->node_spanned_pages) { |
| 1155 | found = 1; |
| 1156 | break; |
| 1157 | } |
| 1158 | } |
| 1159 | |
| 1160 | BUG_ON(!found); |
| 1161 | return nid; |
| 1162 | } |
| 1163 | |
| 1164 | static u64 hot_add_drconf_memory_max(void) |
| 1165 | { |
| 1166 | struct device_node *memory = NULL; |
| 1167 | unsigned int drconf_cell_cnt = 0; |
| 1168 | u64 lmb_size = 0; |
| 1169 | const __be32 *dm = NULL; |
| 1170 | |
| 1171 | memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); |
| 1172 | if (memory) { |
| 1173 | drconf_cell_cnt = of_get_drconf_memory(memory, &dm); |
| 1174 | lmb_size = of_get_lmb_size(memory); |
| 1175 | of_node_put(memory); |
| 1176 | } |
| 1177 | return lmb_size * drconf_cell_cnt; |
| 1178 | } |
| 1179 | |
| 1180 | /* |
| 1181 | * memory_hotplug_max - return max address of memory that may be added |
| 1182 | * |
| 1183 | * This is currently only used on systems that support drconfig memory |
| 1184 | * hotplug. |
| 1185 | */ |
| 1186 | u64 memory_hotplug_max(void) |
| 1187 | { |
| 1188 | return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM()); |
| 1189 | } |
| 1190 | #endif /* CONFIG_MEMORY_HOTPLUG */ |
| 1191 | |
| 1192 | /* Virtual Processor Home Node (VPHN) support */ |
| 1193 | #ifdef CONFIG_PPC_SPLPAR |
| 1194 | |
| 1195 | #include "vphn.h" |
| 1196 | |
| 1197 | struct topology_update_data { |
| 1198 | struct topology_update_data *next; |
| 1199 | unsigned int cpu; |
| 1200 | int old_nid; |
| 1201 | int new_nid; |
| 1202 | }; |
| 1203 | |
| 1204 | static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS]; |
| 1205 | static cpumask_t cpu_associativity_changes_mask; |
| 1206 | static int vphn_enabled; |
| 1207 | static int prrn_enabled; |
| 1208 | static void reset_topology_timer(void); |
| 1209 | |
| 1210 | /* |
| 1211 | * Store the current values of the associativity change counters in the |
| 1212 | * hypervisor. |
| 1213 | */ |
| 1214 | static void setup_cpu_associativity_change_counters(void) |
| 1215 | { |
| 1216 | int cpu; |
| 1217 | |
| 1218 | /* The VPHN feature supports a maximum of 8 reference points */ |
| 1219 | BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8); |
| 1220 | |
| 1221 | for_each_possible_cpu(cpu) { |
| 1222 | int i; |
| 1223 | u8 *counts = vphn_cpu_change_counts[cpu]; |
| 1224 | volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts; |
| 1225 | |
| 1226 | for (i = 0; i < distance_ref_points_depth; i++) |
| 1227 | counts[i] = hypervisor_counts[i]; |
| 1228 | } |
| 1229 | } |
| 1230 | |
| 1231 | /* |
| 1232 | * The hypervisor maintains a set of 8 associativity change counters in |
| 1233 | * the VPA of each cpu that correspond to the associativity levels in the |
| 1234 | * ibm,associativity-reference-points property. When an associativity |
| 1235 | * level changes, the corresponding counter is incremented. |
| 1236 | * |
| 1237 | * Set a bit in cpu_associativity_changes_mask for each cpu whose home |
| 1238 | * node associativity levels have changed. |
| 1239 | * |
| 1240 | * Returns the number of cpus with unhandled associativity changes. |
| 1241 | */ |
| 1242 | static int update_cpu_associativity_changes_mask(void) |
| 1243 | { |
| 1244 | int cpu; |
| 1245 | cpumask_t *changes = &cpu_associativity_changes_mask; |
| 1246 | |
| 1247 | for_each_possible_cpu(cpu) { |
| 1248 | int i, changed = 0; |
| 1249 | u8 *counts = vphn_cpu_change_counts[cpu]; |
| 1250 | volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts; |
| 1251 | |
| 1252 | for (i = 0; i < distance_ref_points_depth; i++) { |
| 1253 | if (hypervisor_counts[i] != counts[i]) { |
| 1254 | counts[i] = hypervisor_counts[i]; |
| 1255 | changed = 1; |
| 1256 | } |
| 1257 | } |
| 1258 | if (changed) { |
| 1259 | cpumask_or(changes, changes, cpu_sibling_mask(cpu)); |
| 1260 | cpu = cpu_last_thread_sibling(cpu); |
| 1261 | } |
| 1262 | } |
| 1263 | |
| 1264 | return cpumask_weight(changes); |
| 1265 | } |
| 1266 | |
| 1267 | /* |
| 1268 | * Retrieve the new associativity information for a virtual processor's |
| 1269 | * home node. |
| 1270 | */ |
| 1271 | static long hcall_vphn(unsigned long cpu, __be32 *associativity) |
| 1272 | { |
| 1273 | long rc; |
| 1274 | long retbuf[PLPAR_HCALL9_BUFSIZE] = {0}; |
| 1275 | u64 flags = 1; |
| 1276 | int hwcpu = get_hard_smp_processor_id(cpu); |
| 1277 | |
| 1278 | rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu); |
| 1279 | vphn_unpack_associativity(retbuf, associativity); |
| 1280 | |
| 1281 | return rc; |
| 1282 | } |
| 1283 | |
| 1284 | static long vphn_get_associativity(unsigned long cpu, |
| 1285 | __be32 *associativity) |
| 1286 | { |
| 1287 | long rc; |
| 1288 | |
| 1289 | rc = hcall_vphn(cpu, associativity); |
| 1290 | |
| 1291 | switch (rc) { |
| 1292 | case H_FUNCTION: |
| 1293 | printk(KERN_INFO |
| 1294 | "VPHN is not supported. Disabling polling...\n"); |
| 1295 | stop_topology_update(); |
| 1296 | break; |
| 1297 | case H_HARDWARE: |
| 1298 | printk(KERN_ERR |
| 1299 | "hcall_vphn() experienced a hardware fault " |
| 1300 | "preventing VPHN. Disabling polling...\n"); |
| 1301 | stop_topology_update(); |
| 1302 | } |
| 1303 | |
| 1304 | return rc; |
| 1305 | } |
| 1306 | |
| 1307 | /* |
| 1308 | * Update the CPU maps and sysfs entries for a single CPU when its NUMA |
| 1309 | * characteristics change. This function doesn't perform any locking and is |
| 1310 | * only safe to call from stop_machine(). |
| 1311 | */ |
| 1312 | static int update_cpu_topology(void *data) |
| 1313 | { |
| 1314 | struct topology_update_data *update; |
| 1315 | unsigned long cpu; |
| 1316 | |
| 1317 | if (!data) |
| 1318 | return -EINVAL; |
| 1319 | |
| 1320 | cpu = smp_processor_id(); |
| 1321 | |
| 1322 | for (update = data; update; update = update->next) { |
| 1323 | int new_nid = update->new_nid; |
| 1324 | if (cpu != update->cpu) |
| 1325 | continue; |
| 1326 | |
| 1327 | unmap_cpu_from_node(cpu); |
| 1328 | map_cpu_to_node(cpu, new_nid); |
| 1329 | set_cpu_numa_node(cpu, new_nid); |
| 1330 | set_cpu_numa_mem(cpu, local_memory_node(new_nid)); |
| 1331 | vdso_getcpu_init(); |
| 1332 | } |
| 1333 | |
| 1334 | return 0; |
| 1335 | } |
| 1336 | |
| 1337 | static int update_lookup_table(void *data) |
| 1338 | { |
| 1339 | struct topology_update_data *update; |
| 1340 | |
| 1341 | if (!data) |
| 1342 | return -EINVAL; |
| 1343 | |
| 1344 | /* |
| 1345 | * Upon topology update, the numa-cpu lookup table needs to be updated |
| 1346 | * for all threads in the core, including offline CPUs, to ensure that |
| 1347 | * future hotplug operations respect the cpu-to-node associativity |
| 1348 | * properly. |
| 1349 | */ |
| 1350 | for (update = data; update; update = update->next) { |
| 1351 | int nid, base, j; |
| 1352 | |
| 1353 | nid = update->new_nid; |
| 1354 | base = cpu_first_thread_sibling(update->cpu); |
| 1355 | |
| 1356 | for (j = 0; j < threads_per_core; j++) { |
| 1357 | update_numa_cpu_lookup_table(base + j, nid); |
| 1358 | } |
| 1359 | } |
| 1360 | |
| 1361 | return 0; |
| 1362 | } |
| 1363 | |
| 1364 | /* |
| 1365 | * Update the node maps and sysfs entries for each cpu whose home node |
| 1366 | * has changed. Returns 1 when the topology has changed, and 0 otherwise. |
| 1367 | */ |
| 1368 | int arch_update_cpu_topology(void) |
| 1369 | { |
| 1370 | unsigned int cpu, sibling, changed = 0; |
| 1371 | struct topology_update_data *updates, *ud; |
| 1372 | __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0}; |
| 1373 | cpumask_t updated_cpus; |
| 1374 | struct device *dev; |
| 1375 | int weight, new_nid, i = 0; |
| 1376 | |
| 1377 | if (!prrn_enabled && !vphn_enabled) |
| 1378 | return 0; |
| 1379 | |
| 1380 | weight = cpumask_weight(&cpu_associativity_changes_mask); |
| 1381 | if (!weight) |
| 1382 | return 0; |
| 1383 | |
| 1384 | updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL); |
| 1385 | if (!updates) |
| 1386 | return 0; |
| 1387 | |
| 1388 | cpumask_clear(&updated_cpus); |
| 1389 | |
| 1390 | for_each_cpu(cpu, &cpu_associativity_changes_mask) { |
| 1391 | /* |
| 1392 | * If siblings aren't flagged for changes, updates list |
| 1393 | * will be too short. Skip on this update and set for next |
| 1394 | * update. |
| 1395 | */ |
| 1396 | if (!cpumask_subset(cpu_sibling_mask(cpu), |
| 1397 | &cpu_associativity_changes_mask)) { |
| 1398 | pr_info("Sibling bits not set for associativity " |
| 1399 | "change, cpu%d\n", cpu); |
| 1400 | cpumask_or(&cpu_associativity_changes_mask, |
| 1401 | &cpu_associativity_changes_mask, |
| 1402 | cpu_sibling_mask(cpu)); |
| 1403 | cpu = cpu_last_thread_sibling(cpu); |
| 1404 | continue; |
| 1405 | } |
| 1406 | |
| 1407 | /* Use associativity from first thread for all siblings */ |
| 1408 | vphn_get_associativity(cpu, associativity); |
| 1409 | new_nid = associativity_to_nid(associativity); |
| 1410 | if (new_nid < 0 || !node_online(new_nid)) |
| 1411 | new_nid = first_online_node; |
| 1412 | |
| 1413 | if (new_nid == numa_cpu_lookup_table[cpu]) { |
| 1414 | cpumask_andnot(&cpu_associativity_changes_mask, |
| 1415 | &cpu_associativity_changes_mask, |
| 1416 | cpu_sibling_mask(cpu)); |
| 1417 | cpu = cpu_last_thread_sibling(cpu); |
| 1418 | continue; |
| 1419 | } |
| 1420 | |
| 1421 | for_each_cpu(sibling, cpu_sibling_mask(cpu)) { |
| 1422 | ud = &updates[i++]; |
| 1423 | ud->cpu = sibling; |
| 1424 | ud->new_nid = new_nid; |
| 1425 | ud->old_nid = numa_cpu_lookup_table[sibling]; |
| 1426 | cpumask_set_cpu(sibling, &updated_cpus); |
| 1427 | if (i < weight) |
| 1428 | ud->next = &updates[i]; |
| 1429 | } |
| 1430 | cpu = cpu_last_thread_sibling(cpu); |
| 1431 | } |
| 1432 | |
| 1433 | pr_debug("Topology update for the following CPUs:\n"); |
| 1434 | if (cpumask_weight(&updated_cpus)) { |
| 1435 | for (ud = &updates[0]; ud; ud = ud->next) { |
| 1436 | pr_debug("cpu %d moving from node %d " |
| 1437 | "to %d\n", ud->cpu, |
| 1438 | ud->old_nid, ud->new_nid); |
| 1439 | } |
| 1440 | } |
| 1441 | |
| 1442 | /* |
| 1443 | * In cases where we have nothing to update (because the updates list |
| 1444 | * is too short or because the new topology is same as the old one), |
| 1445 | * skip invoking update_cpu_topology() via stop-machine(). This is |
| 1446 | * necessary (and not just a fast-path optimization) since stop-machine |
| 1447 | * can end up electing a random CPU to run update_cpu_topology(), and |
| 1448 | * thus trick us into setting up incorrect cpu-node mappings (since |
| 1449 | * 'updates' is kzalloc()'ed). |
| 1450 | * |
| 1451 | * And for the similar reason, we will skip all the following updating. |
| 1452 | */ |
| 1453 | if (!cpumask_weight(&updated_cpus)) |
| 1454 | goto out; |
| 1455 | |
| 1456 | stop_machine(update_cpu_topology, &updates[0], &updated_cpus); |
| 1457 | |
| 1458 | /* |
| 1459 | * Update the numa-cpu lookup table with the new mappings, even for |
| 1460 | * offline CPUs. It is best to perform this update from the stop- |
| 1461 | * machine context. |
| 1462 | */ |
| 1463 | stop_machine(update_lookup_table, &updates[0], |
| 1464 | cpumask_of(raw_smp_processor_id())); |
| 1465 | |
| 1466 | for (ud = &updates[0]; ud; ud = ud->next) { |
| 1467 | unregister_cpu_under_node(ud->cpu, ud->old_nid); |
| 1468 | register_cpu_under_node(ud->cpu, ud->new_nid); |
| 1469 | |
| 1470 | dev = get_cpu_device(ud->cpu); |
| 1471 | if (dev) |
| 1472 | kobject_uevent(&dev->kobj, KOBJ_CHANGE); |
| 1473 | cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask); |
| 1474 | changed = 1; |
| 1475 | } |
| 1476 | |
| 1477 | out: |
| 1478 | kfree(updates); |
| 1479 | return changed; |
| 1480 | } |
| 1481 | |
| 1482 | static void topology_work_fn(struct work_struct *work) |
| 1483 | { |
| 1484 | rebuild_sched_domains(); |
| 1485 | } |
| 1486 | static DECLARE_WORK(topology_work, topology_work_fn); |
| 1487 | |
| 1488 | static void topology_schedule_update(void) |
| 1489 | { |
| 1490 | schedule_work(&topology_work); |
| 1491 | } |
| 1492 | |
| 1493 | static void topology_timer_fn(unsigned long ignored) |
| 1494 | { |
| 1495 | if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask)) |
| 1496 | topology_schedule_update(); |
| 1497 | else if (vphn_enabled) { |
| 1498 | if (update_cpu_associativity_changes_mask() > 0) |
| 1499 | topology_schedule_update(); |
| 1500 | reset_topology_timer(); |
| 1501 | } |
| 1502 | } |
| 1503 | static struct timer_list topology_timer = |
| 1504 | TIMER_INITIALIZER(topology_timer_fn, 0, 0); |
| 1505 | |
| 1506 | static void reset_topology_timer(void) |
| 1507 | { |
| 1508 | topology_timer.data = 0; |
| 1509 | topology_timer.expires = jiffies + 60 * HZ; |
| 1510 | mod_timer(&topology_timer, topology_timer.expires); |
| 1511 | } |
| 1512 | |
| 1513 | #ifdef CONFIG_SMP |
| 1514 | |
| 1515 | static void stage_topology_update(int core_id) |
| 1516 | { |
| 1517 | cpumask_or(&cpu_associativity_changes_mask, |
| 1518 | &cpu_associativity_changes_mask, cpu_sibling_mask(core_id)); |
| 1519 | reset_topology_timer(); |
| 1520 | } |
| 1521 | |
| 1522 | static int dt_update_callback(struct notifier_block *nb, |
| 1523 | unsigned long action, void *data) |
| 1524 | { |
| 1525 | struct of_reconfig_data *update = data; |
| 1526 | int rc = NOTIFY_DONE; |
| 1527 | |
| 1528 | switch (action) { |
| 1529 | case OF_RECONFIG_UPDATE_PROPERTY: |
| 1530 | if (!of_prop_cmp(update->dn->type, "cpu") && |
| 1531 | !of_prop_cmp(update->prop->name, "ibm,associativity")) { |
| 1532 | u32 core_id; |
| 1533 | of_property_read_u32(update->dn, "reg", &core_id); |
| 1534 | stage_topology_update(core_id); |
| 1535 | rc = NOTIFY_OK; |
| 1536 | } |
| 1537 | break; |
| 1538 | } |
| 1539 | |
| 1540 | return rc; |
| 1541 | } |
| 1542 | |
| 1543 | static struct notifier_block dt_update_nb = { |
| 1544 | .notifier_call = dt_update_callback, |
| 1545 | }; |
| 1546 | |
| 1547 | #endif |
| 1548 | |
| 1549 | /* |
| 1550 | * Start polling for associativity changes. |
| 1551 | */ |
| 1552 | int start_topology_update(void) |
| 1553 | { |
| 1554 | int rc = 0; |
| 1555 | |
| 1556 | if (firmware_has_feature(FW_FEATURE_PRRN)) { |
| 1557 | if (!prrn_enabled) { |
| 1558 | prrn_enabled = 1; |
| 1559 | vphn_enabled = 0; |
| 1560 | #ifdef CONFIG_SMP |
| 1561 | rc = of_reconfig_notifier_register(&dt_update_nb); |
| 1562 | #endif |
| 1563 | } |
| 1564 | } else if (firmware_has_feature(FW_FEATURE_VPHN) && |
| 1565 | lppaca_shared_proc(get_lppaca())) { |
| 1566 | if (!vphn_enabled) { |
| 1567 | prrn_enabled = 0; |
| 1568 | vphn_enabled = 1; |
| 1569 | setup_cpu_associativity_change_counters(); |
| 1570 | init_timer_deferrable(&topology_timer); |
| 1571 | reset_topology_timer(); |
| 1572 | } |
| 1573 | } |
| 1574 | |
| 1575 | return rc; |
| 1576 | } |
| 1577 | |
| 1578 | /* |
| 1579 | * Disable polling for VPHN associativity changes. |
| 1580 | */ |
| 1581 | int stop_topology_update(void) |
| 1582 | { |
| 1583 | int rc = 0; |
| 1584 | |
| 1585 | if (prrn_enabled) { |
| 1586 | prrn_enabled = 0; |
| 1587 | #ifdef CONFIG_SMP |
| 1588 | rc = of_reconfig_notifier_unregister(&dt_update_nb); |
| 1589 | #endif |
| 1590 | } else if (vphn_enabled) { |
| 1591 | vphn_enabled = 0; |
| 1592 | rc = del_timer_sync(&topology_timer); |
| 1593 | } |
| 1594 | |
| 1595 | return rc; |
| 1596 | } |
| 1597 | |
| 1598 | int prrn_is_enabled(void) |
| 1599 | { |
| 1600 | return prrn_enabled; |
| 1601 | } |
| 1602 | |
| 1603 | static int topology_read(struct seq_file *file, void *v) |
| 1604 | { |
| 1605 | if (vphn_enabled || prrn_enabled) |
| 1606 | seq_puts(file, "on\n"); |
| 1607 | else |
| 1608 | seq_puts(file, "off\n"); |
| 1609 | |
| 1610 | return 0; |
| 1611 | } |
| 1612 | |
| 1613 | static int topology_open(struct inode *inode, struct file *file) |
| 1614 | { |
| 1615 | return single_open(file, topology_read, NULL); |
| 1616 | } |
| 1617 | |
| 1618 | static ssize_t topology_write(struct file *file, const char __user *buf, |
| 1619 | size_t count, loff_t *off) |
| 1620 | { |
| 1621 | char kbuf[4]; /* "on" or "off" plus null. */ |
| 1622 | int read_len; |
| 1623 | |
| 1624 | read_len = count < 3 ? count : 3; |
| 1625 | if (copy_from_user(kbuf, buf, read_len)) |
| 1626 | return -EINVAL; |
| 1627 | |
| 1628 | kbuf[read_len] = '\0'; |
| 1629 | |
| 1630 | if (!strncmp(kbuf, "on", 2)) |
| 1631 | start_topology_update(); |
| 1632 | else if (!strncmp(kbuf, "off", 3)) |
| 1633 | stop_topology_update(); |
| 1634 | else |
| 1635 | return -EINVAL; |
| 1636 | |
| 1637 | return count; |
| 1638 | } |
| 1639 | |
| 1640 | static const struct file_operations topology_ops = { |
| 1641 | .read = seq_read, |
| 1642 | .write = topology_write, |
| 1643 | .open = topology_open, |
| 1644 | .release = single_release |
| 1645 | }; |
| 1646 | |
| 1647 | static int topology_update_init(void) |
| 1648 | { |
| 1649 | /* Do not poll for changes if disabled at boot */ |
| 1650 | if (topology_updates_enabled) |
| 1651 | start_topology_update(); |
| 1652 | |
| 1653 | if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops)) |
| 1654 | return -ENOMEM; |
| 1655 | |
| 1656 | return 0; |
| 1657 | } |
| 1658 | device_initcall(topology_update_init); |
| 1659 | #endif /* CONFIG_PPC_SPLPAR */ |