linux/arch/x86/mm/numa_emulation.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * NUMA emulation
   4 */
   5#include <linux/kernel.h>
   6#include <linux/errno.h>
   7#include <linux/topology.h>
   8#include <linux/memblock.h>
   9#include <linux/bootmem.h>
  10#include <asm/dma.h>
  11
  12#include "numa_internal.h"
  13
  14static int emu_nid_to_phys[MAX_NUMNODES];
  15static char *emu_cmdline __initdata;
  16
  17void __init numa_emu_cmdline(char *str)
  18{
  19        emu_cmdline = str;
  20}
  21
  22static int __init emu_find_memblk_by_nid(int nid, const struct numa_meminfo *mi)
  23{
  24        int i;
  25
  26        for (i = 0; i < mi->nr_blks; i++)
  27                if (mi->blk[i].nid == nid)
  28                        return i;
  29        return -ENOENT;
  30}
  31
  32static u64 __init mem_hole_size(u64 start, u64 end)
  33{
  34        unsigned long start_pfn = PFN_UP(start);
  35        unsigned long end_pfn = PFN_DOWN(end);
  36
  37        if (start_pfn < end_pfn)
  38                return PFN_PHYS(absent_pages_in_range(start_pfn, end_pfn));
  39        return 0;
  40}
  41
  42/*
  43 * Sets up nid to range from @start to @end.  The return value is -errno if
  44 * something went wrong, 0 otherwise.
  45 */
  46static int __init emu_setup_memblk(struct numa_meminfo *ei,
  47                                   struct numa_meminfo *pi,
  48                                   int nid, int phys_blk, u64 size)
  49{
  50        struct numa_memblk *eb = &ei->blk[ei->nr_blks];
  51        struct numa_memblk *pb = &pi->blk[phys_blk];
  52
  53        if (ei->nr_blks >= NR_NODE_MEMBLKS) {
  54                pr_err("NUMA: Too many emulated memblks, failing emulation\n");
  55                return -EINVAL;
  56        }
  57
  58        ei->nr_blks++;
  59        eb->start = pb->start;
  60        eb->end = pb->start + size;
  61        eb->nid = nid;
  62
  63        if (emu_nid_to_phys[nid] == NUMA_NO_NODE)
  64                emu_nid_to_phys[nid] = nid;
  65
  66        pb->start += size;
  67        if (pb->start >= pb->end) {
  68                WARN_ON_ONCE(pb->start > pb->end);
  69                numa_remove_memblk_from(phys_blk, pi);
  70        }
  71
  72        printk(KERN_INFO "Faking node %d at [mem %#018Lx-%#018Lx] (%LuMB)\n",
  73               nid, eb->start, eb->end - 1, (eb->end - eb->start) >> 20);
  74        return 0;
  75}
  76
  77/*
  78 * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr
  79 * to max_addr.
  80 *
  81 * Returns zero on success or negative on error.
  82 */
  83static int __init split_nodes_interleave(struct numa_meminfo *ei,
  84                                         struct numa_meminfo *pi,
  85                                         u64 addr, u64 max_addr, int nr_nodes)
  86{
  87        nodemask_t physnode_mask = numa_nodes_parsed;
  88        u64 size;
  89        int big;
  90        int nid = 0;
  91        int i, ret;
  92
  93        if (nr_nodes <= 0)
  94                return -1;
  95        if (nr_nodes > MAX_NUMNODES) {
  96                pr_info("numa=fake=%d too large, reducing to %d\n",
  97                        nr_nodes, MAX_NUMNODES);
  98                nr_nodes = MAX_NUMNODES;
  99        }
 100
 101        /*
 102         * Calculate target node size.  x86_32 freaks on __udivdi3() so do
 103         * the division in ulong number of pages and convert back.
 104         */
 105        size = max_addr - addr - mem_hole_size(addr, max_addr);
 106        size = PFN_PHYS((unsigned long)(size >> PAGE_SHIFT) / nr_nodes);
 107
 108        /*
 109         * Calculate the number of big nodes that can be allocated as a result
 110         * of consolidating the remainder.
 111         */
 112        big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * nr_nodes) /
 113                FAKE_NODE_MIN_SIZE;
 114
 115        size &= FAKE_NODE_MIN_HASH_MASK;
 116        if (!size) {
 117                pr_err("Not enough memory for each node.  "
 118                        "NUMA emulation disabled.\n");
 119                return -1;
 120        }
 121
 122        /*
 123         * Continue to fill physical nodes with fake nodes until there is no
 124         * memory left on any of them.
 125         */
 126        while (nodes_weight(physnode_mask)) {
 127                for_each_node_mask(i, physnode_mask) {
 128                        u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
 129                        u64 start, limit, end;
 130                        int phys_blk;
 131
 132                        phys_blk = emu_find_memblk_by_nid(i, pi);
 133                        if (phys_blk < 0) {
 134                                node_clear(i, physnode_mask);
 135                                continue;
 136                        }
 137                        start = pi->blk[phys_blk].start;
 138                        limit = pi->blk[phys_blk].end;
 139                        end = start + size;
 140
 141                        if (nid < big)
 142                                end += FAKE_NODE_MIN_SIZE;
 143
 144                        /*
 145                         * Continue to add memory to this fake node if its
 146                         * non-reserved memory is less than the per-node size.
 147                         */
 148                        while (end - start - mem_hole_size(start, end) < size) {
 149                                end += FAKE_NODE_MIN_SIZE;
 150                                if (end > limit) {
 151                                        end = limit;
 152                                        break;
 153                                }
 154                        }
 155
 156                        /*
 157                         * If there won't be at least FAKE_NODE_MIN_SIZE of
 158                         * non-reserved memory in ZONE_DMA32 for the next node,
 159                         * this one must extend to the boundary.
 160                         */
 161                        if (end < dma32_end && dma32_end - end -
 162                            mem_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
 163                                end = dma32_end;
 164
 165                        /*
 166                         * If there won't be enough non-reserved memory for the
 167                         * next node, this one must extend to the end of the
 168                         * physical node.
 169                         */
 170                        if (limit - end - mem_hole_size(end, limit) < size)
 171                                end = limit;
 172
 173                        ret = emu_setup_memblk(ei, pi, nid++ % nr_nodes,
 174                                               phys_blk,
 175                                               min(end, limit) - start);
 176                        if (ret < 0)
 177                                return ret;
 178                }
 179        }
 180        return 0;
 181}
 182
 183/*
 184 * Returns the end address of a node so that there is at least `size' amount of
 185 * non-reserved memory or `max_addr' is reached.
 186 */
 187static u64 __init find_end_of_node(u64 start, u64 max_addr, u64 size)
 188{
 189        u64 end = start + size;
 190
 191        while (end - start - mem_hole_size(start, end) < size) {
 192                end += FAKE_NODE_MIN_SIZE;
 193                if (end > max_addr) {
 194                        end = max_addr;
 195                        break;
 196                }
 197        }
 198        return end;
 199}
 200
 201/*
 202 * Sets up fake nodes of `size' interleaved over physical nodes ranging from
 203 * `addr' to `max_addr'.
 204 *
 205 * Returns zero on success or negative on error.
 206 */
 207static int __init split_nodes_size_interleave(struct numa_meminfo *ei,
 208                                              struct numa_meminfo *pi,
 209                                              u64 addr, u64 max_addr, u64 size)
 210{
 211        nodemask_t physnode_mask = numa_nodes_parsed;
 212        u64 min_size;
 213        int nid = 0;
 214        int i, ret;
 215
 216        if (!size)
 217                return -1;
 218        /*
 219         * The limit on emulated nodes is MAX_NUMNODES, so the size per node is
 220         * increased accordingly if the requested size is too small.  This
 221         * creates a uniform distribution of node sizes across the entire
 222         * machine (but not necessarily over physical nodes).
 223         */
 224        min_size = (max_addr - addr - mem_hole_size(addr, max_addr)) / MAX_NUMNODES;
 225        min_size = max(min_size, FAKE_NODE_MIN_SIZE);
 226        if ((min_size & FAKE_NODE_MIN_HASH_MASK) < min_size)
 227                min_size = (min_size + FAKE_NODE_MIN_SIZE) &
 228                                                FAKE_NODE_MIN_HASH_MASK;
 229        if (size < min_size) {
 230                pr_err("Fake node size %LuMB too small, increasing to %LuMB\n",
 231                        size >> 20, min_size >> 20);
 232                size = min_size;
 233        }
 234        size &= FAKE_NODE_MIN_HASH_MASK;
 235
 236        /*
 237         * Fill physical nodes with fake nodes of size until there is no memory
 238         * left on any of them.
 239         */
 240        while (nodes_weight(physnode_mask)) {
 241                for_each_node_mask(i, physnode_mask) {
 242                        u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
 243                        u64 start, limit, end;
 244                        int phys_blk;
 245
 246                        phys_blk = emu_find_memblk_by_nid(i, pi);
 247                        if (phys_blk < 0) {
 248                                node_clear(i, physnode_mask);
 249                                continue;
 250                        }
 251                        start = pi->blk[phys_blk].start;
 252                        limit = pi->blk[phys_blk].end;
 253
 254                        end = find_end_of_node(start, limit, size);
 255                        /*
 256                         * If there won't be at least FAKE_NODE_MIN_SIZE of
 257                         * non-reserved memory in ZONE_DMA32 for the next node,
 258                         * this one must extend to the boundary.
 259                         */
 260                        if (end < dma32_end && dma32_end - end -
 261                            mem_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
 262                                end = dma32_end;
 263
 264                        /*
 265                         * If there won't be enough non-reserved memory for the
 266                         * next node, this one must extend to the end of the
 267                         * physical node.
 268                         */
 269                        if (limit - end - mem_hole_size(end, limit) < size)
 270                                end = limit;
 271
 272                        ret = emu_setup_memblk(ei, pi, nid++ % MAX_NUMNODES,
 273                                               phys_blk,
 274                                               min(end, limit) - start);
 275                        if (ret < 0)
 276                                return ret;
 277                }
 278        }
 279        return 0;
 280}
 281
 282int __init setup_emu2phys_nid(int *dfl_phys_nid)
 283{
 284        int i, max_emu_nid = 0;
 285
 286        *dfl_phys_nid = NUMA_NO_NODE;
 287        for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++) {
 288                if (emu_nid_to_phys[i] != NUMA_NO_NODE) {
 289                        max_emu_nid = i;
 290                        if (*dfl_phys_nid == NUMA_NO_NODE)
 291                                *dfl_phys_nid = emu_nid_to_phys[i];
 292                }
 293        }
 294
 295        return max_emu_nid;
 296}
 297
 298/**
 299 * numa_emulation - Emulate NUMA nodes
 300 * @numa_meminfo: NUMA configuration to massage
 301 * @numa_dist_cnt: The size of the physical NUMA distance table
 302 *
 303 * Emulate NUMA nodes according to the numa=fake kernel parameter.
 304 * @numa_meminfo contains the physical memory configuration and is modified
 305 * to reflect the emulated configuration on success.  @numa_dist_cnt is
 306 * used to determine the size of the physical distance table.
 307 *
 308 * On success, the following modifications are made.
 309 *
 310 * - @numa_meminfo is updated to reflect the emulated nodes.
 311 *
 312 * - __apicid_to_node[] is updated such that APIC IDs are mapped to the
 313 *   emulated nodes.
 314 *
 315 * - NUMA distance table is rebuilt to represent distances between emulated
 316 *   nodes.  The distances are determined considering how emulated nodes
 317 *   are mapped to physical nodes and match the actual distances.
 318 *
 319 * - emu_nid_to_phys[] reflects how emulated nodes are mapped to physical
 320 *   nodes.  This is used by numa_add_cpu() and numa_remove_cpu().
 321 *
 322 * If emulation is not enabled or fails, emu_nid_to_phys[] is filled with
 323 * identity mapping and no other modification is made.
 324 */
 325void __init numa_emulation(struct numa_meminfo *numa_meminfo, int numa_dist_cnt)
 326{
 327        static struct numa_meminfo ei __initdata;
 328        static struct numa_meminfo pi __initdata;
 329        const u64 max_addr = PFN_PHYS(max_pfn);
 330        u8 *phys_dist = NULL;
 331        size_t phys_size = numa_dist_cnt * numa_dist_cnt * sizeof(phys_dist[0]);
 332        int max_emu_nid, dfl_phys_nid;
 333        int i, j, ret;
 334
 335        if (!emu_cmdline)
 336                goto no_emu;
 337
 338        memset(&ei, 0, sizeof(ei));
 339        pi = *numa_meminfo;
 340
 341        for (i = 0; i < MAX_NUMNODES; i++)
 342                emu_nid_to_phys[i] = NUMA_NO_NODE;
 343
 344        /*
 345         * If the numa=fake command-line contains a 'M' or 'G', it represents
 346         * the fixed node size.  Otherwise, if it is just a single number N,
 347         * split the system RAM into N fake nodes.
 348         */
 349        if (strchr(emu_cmdline, 'M') || strchr(emu_cmdline, 'G')) {
 350                u64 size;
 351
 352                size = memparse(emu_cmdline, &emu_cmdline);
 353                ret = split_nodes_size_interleave(&ei, &pi, 0, max_addr, size);
 354        } else {
 355                unsigned long n;
 356
 357                n = simple_strtoul(emu_cmdline, &emu_cmdline, 0);
 358                ret = split_nodes_interleave(&ei, &pi, 0, max_addr, n);
 359        }
 360        if (*emu_cmdline == ':')
 361                emu_cmdline++;
 362
 363        if (ret < 0)
 364                goto no_emu;
 365
 366        if (numa_cleanup_meminfo(&ei) < 0) {
 367                pr_warning("NUMA: Warning: constructed meminfo invalid, disabling emulation\n");
 368                goto no_emu;
 369        }
 370
 371        /* copy the physical distance table */
 372        if (numa_dist_cnt) {
 373                u64 phys;
 374
 375                phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
 376                                              phys_size, PAGE_SIZE);
 377                if (!phys) {
 378                        pr_warning("NUMA: Warning: can't allocate copy of distance table, disabling emulation\n");
 379                        goto no_emu;
 380                }
 381                memblock_reserve(phys, phys_size);
 382                phys_dist = __va(phys);
 383
 384                for (i = 0; i < numa_dist_cnt; i++)
 385                        for (j = 0; j < numa_dist_cnt; j++)
 386                                phys_dist[i * numa_dist_cnt + j] =
 387                                        node_distance(i, j);
 388        }
 389
 390        /*
 391         * Determine the max emulated nid and the default phys nid to use
 392         * for unmapped nodes.
 393         */
 394        max_emu_nid = setup_emu2phys_nid(&dfl_phys_nid);
 395
 396        /* commit */
 397        *numa_meminfo = ei;
 398
 399        /* Make sure numa_nodes_parsed only contains emulated nodes */
 400        nodes_clear(numa_nodes_parsed);
 401        for (i = 0; i < ARRAY_SIZE(ei.blk); i++)
 402                if (ei.blk[i].start != ei.blk[i].end &&
 403                    ei.blk[i].nid != NUMA_NO_NODE)
 404                        node_set(ei.blk[i].nid, numa_nodes_parsed);
 405
 406        /*
 407         * Transform __apicid_to_node table to use emulated nids by
 408         * reverse-mapping phys_nid.  The maps should always exist but fall
 409         * back to zero just in case.
 410         */
 411        for (i = 0; i < ARRAY_SIZE(__apicid_to_node); i++) {
 412                if (__apicid_to_node[i] == NUMA_NO_NODE)
 413                        continue;
 414                for (j = 0; j < ARRAY_SIZE(emu_nid_to_phys); j++)
 415                        if (__apicid_to_node[i] == emu_nid_to_phys[j])
 416                                break;
 417                __apicid_to_node[i] = j < ARRAY_SIZE(emu_nid_to_phys) ? j : 0;
 418        }
 419
 420        /* make sure all emulated nodes are mapped to a physical node */
 421        for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++)
 422                if (emu_nid_to_phys[i] == NUMA_NO_NODE)
 423                        emu_nid_to_phys[i] = dfl_phys_nid;
 424
 425        /* transform distance table */
 426        numa_reset_distance();
 427        for (i = 0; i < max_emu_nid + 1; i++) {
 428                for (j = 0; j < max_emu_nid + 1; j++) {
 429                        int physi = emu_nid_to_phys[i];
 430                        int physj = emu_nid_to_phys[j];
 431                        int dist;
 432
 433                        if (get_option(&emu_cmdline, &dist) == 2)
 434                                ;
 435                        else if (physi >= numa_dist_cnt || physj >= numa_dist_cnt)
 436                                dist = physi == physj ?
 437                                        LOCAL_DISTANCE : REMOTE_DISTANCE;
 438                        else
 439                                dist = phys_dist[physi * numa_dist_cnt + physj];
 440
 441                        numa_set_distance(i, j, dist);
 442                }
 443        }
 444
 445        /* free the copied physical distance table */
 446        if (phys_dist)
 447                memblock_free(__pa(phys_dist), phys_size);
 448        return;
 449
 450no_emu:
 451        /* No emulation.  Build identity emu_nid_to_phys[] for numa_add_cpu() */
 452        for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++)
 453                emu_nid_to_phys[i] = i;
 454}
 455
 456#ifndef CONFIG_DEBUG_PER_CPU_MAPS
 457void numa_add_cpu(int cpu)
 458{
 459        int physnid, nid;
 460
 461        nid = early_cpu_to_node(cpu);
 462        BUG_ON(nid == NUMA_NO_NODE || !node_online(nid));
 463
 464        physnid = emu_nid_to_phys[nid];
 465
 466        /*
 467         * Map the cpu to each emulated node that is allocated on the physical
 468         * node of the cpu's apic id.
 469         */
 470        for_each_online_node(nid)
 471                if (emu_nid_to_phys[nid] == physnid)
 472                        cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
 473}
 474
 475void numa_remove_cpu(int cpu)
 476{
 477        int i;
 478
 479        for_each_online_node(i)
 480                cpumask_clear_cpu(cpu, node_to_cpumask_map[i]);
 481}
 482#else   /* !CONFIG_DEBUG_PER_CPU_MAPS */
 483static void numa_set_cpumask(int cpu, bool enable)
 484{
 485        int nid, physnid;
 486
 487        nid = early_cpu_to_node(cpu);
 488        if (nid == NUMA_NO_NODE) {
 489                /* early_cpu_to_node() already emits a warning and trace */
 490                return;
 491        }
 492
 493        physnid = emu_nid_to_phys[nid];
 494
 495        for_each_online_node(nid) {
 496                if (emu_nid_to_phys[nid] != physnid)
 497                        continue;
 498
 499                debug_cpumask_set_cpu(cpu, nid, enable);
 500        }
 501}
 502
 503void numa_add_cpu(int cpu)
 504{
 505        numa_set_cpumask(cpu, true);
 506}
 507
 508void numa_remove_cpu(int cpu)
 509{
 510        numa_set_cpumask(cpu, false);
 511}
 512#endif  /* !CONFIG_DEBUG_PER_CPU_MAPS */
 513