LXR linux/arch/alpha/mm/numa.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  linux/arch/alpha/mm/numa.c
   4 *
   5 *  DISCONTIGMEM NUMA alpha support.
   6 *
   7 *  Copyright (C) 2001 Andrea Arcangeli <andrea@suse.de> SuSE
   8 */
   9
  10#include <linux/types.h>
  11#include <linux/kernel.h>
  12#include <linux/mm.h>
  13#include <linux/bootmem.h>
  14#include <linux/swap.h>
  15#include <linux/initrd.h>
  16#include <linux/pfn.h>
  17#include <linux/module.h>
  18
  19#include <asm/hwrpb.h>
  20#include <asm/pgalloc.h>
  21#include <asm/sections.h>
  22
  23pg_data_t node_data[MAX_NUMNODES];
  24EXPORT_SYMBOL(node_data);
  25
  26#undef DEBUG_DISCONTIG
  27#ifdef DEBUG_DISCONTIG
  28#define DBGDCONT(args...) printk(args)
  29#else
  30#define DBGDCONT(args...)
  31#endif
  32
  33#define for_each_mem_cluster(memdesc, _cluster, i)              \
  34        for ((_cluster) = (memdesc)->cluster, (i) = 0;          \
  35             (i) < (memdesc)->numclusters; (i)++, (_cluster)++)
  36
  37static void __init show_mem_layout(void)
  38{
  39        struct memclust_struct * cluster;
  40        struct memdesc_struct * memdesc;
  41        int i;
  42
  43        /* Find free clusters, and init and free the bootmem accordingly.  */
  44        memdesc = (struct memdesc_struct *)
  45          (hwrpb->mddt_offset + (unsigned long) hwrpb);
  46
  47        printk("Raw memory layout:\n");
  48        for_each_mem_cluster(memdesc, cluster, i) {
  49                printk(" memcluster %2d, usage %1lx, start %8lu, end %8lu\n",
  50                       i, cluster->usage, cluster->start_pfn,
  51                       cluster->start_pfn + cluster->numpages);
  52        }
  53}
  54
  55static void __init
  56setup_memory_node(int nid, void *kernel_end)
  57{
  58        extern unsigned long mem_size_limit;
  59        struct memclust_struct * cluster;
  60        struct memdesc_struct * memdesc;
  61        unsigned long start_kernel_pfn, end_kernel_pfn;
  62        unsigned long bootmap_size, bootmap_pages, bootmap_start;
  63        unsigned long start, end;
  64        unsigned long node_pfn_start, node_pfn_end;
  65        unsigned long node_min_pfn, node_max_pfn;
  66        int i;
  67        unsigned long node_datasz = PFN_UP(sizeof(pg_data_t));
  68        int show_init = 0;
  69
  70        /* Find the bounds of current node */
  71        node_pfn_start = (node_mem_start(nid)) >> PAGE_SHIFT;
  72        node_pfn_end = node_pfn_start + (node_mem_size(nid) >> PAGE_SHIFT);
  73        
  74        /* Find free clusters, and init and free the bootmem accordingly.  */
  75        memdesc = (struct memdesc_struct *)
  76          (hwrpb->mddt_offset + (unsigned long) hwrpb);
  77
  78        /* find the bounds of this node (node_min_pfn/node_max_pfn) */
  79        node_min_pfn = ~0UL;
  80        node_max_pfn = 0UL;
  81        for_each_mem_cluster(memdesc, cluster, i) {
  82                /* Bit 0 is console/PALcode reserved.  Bit 1 is
  83                   non-volatile memory -- we might want to mark
  84                   this for later.  */
  85                if (cluster->usage & 3)
  86                        continue;
  87
  88                start = cluster->start_pfn;
  89                end = start + cluster->numpages;
  90
  91                if (start >= node_pfn_end || end <= node_pfn_start)
  92                        continue;
  93
  94                if (!show_init) {
  95                        show_init = 1;
  96                        printk("Initializing bootmem allocator on Node ID %d\n", nid);
  97                }
  98                printk(" memcluster %2d, usage %1lx, start %8lu, end %8lu\n",
  99                       i, cluster->usage, cluster->start_pfn,
 100                       cluster->start_pfn + cluster->numpages);
 101
 102                if (start < node_pfn_start)
 103                        start = node_pfn_start;
 104                if (end > node_pfn_end)
 105                        end = node_pfn_end;
 106
 107                if (start < node_min_pfn)
 108                        node_min_pfn = start;
 109                if (end > node_max_pfn)
 110                        node_max_pfn = end;
 111        }
 112
 113        if (mem_size_limit && node_max_pfn > mem_size_limit) {
 114                static int msg_shown = 0;
 115                if (!msg_shown) {
 116                        msg_shown = 1;
 117                        printk("setup: forcing memory size to %ldK (from %ldK).\n",
 118                               mem_size_limit << (PAGE_SHIFT - 10),
 119                               node_max_pfn    << (PAGE_SHIFT - 10));
 120                }
 121                node_max_pfn = mem_size_limit;
 122        }
 123
 124        if (node_min_pfn >= node_max_pfn)
 125                return;
 126
 127        /* Update global {min,max}_low_pfn from node information. */
 128        if (node_min_pfn < min_low_pfn)
 129                min_low_pfn = node_min_pfn;
 130        if (node_max_pfn > max_low_pfn)
 131                max_pfn = max_low_pfn = node_max_pfn;
 132
 133#if 0 /* we'll try this one again in a little while */
 134        /* Cute trick to make sure our local node data is on local memory */
 135        node_data[nid] = (pg_data_t *)(__va(node_min_pfn << PAGE_SHIFT));
 136#endif
 137        /* Quasi-mark the pg_data_t as in-use */
 138        node_min_pfn += node_datasz;
 139        if (node_min_pfn >= node_max_pfn) {
 140                printk(" not enough mem to reserve NODE_DATA");
 141                return;
 142        }
 143        NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
 144
 145        printk(" Detected node memory:   start %8lu, end %8lu\n",
 146               node_min_pfn, node_max_pfn);
 147
 148        DBGDCONT(" DISCONTIG: node_data[%d]   is at 0x%p\n", nid, NODE_DATA(nid));
 149        DBGDCONT(" DISCONTIG: NODE_DATA(%d)->bdata is at 0x%p\n", nid, NODE_DATA(nid)->bdata);
 150
 151        /* Find the bounds of kernel memory.  */
 152        start_kernel_pfn = PFN_DOWN(KERNEL_START_PHYS);
 153        end_kernel_pfn = PFN_UP(virt_to_phys(kernel_end));
 154        bootmap_start = -1;
 155
 156        if (!nid && (node_max_pfn < end_kernel_pfn || node_min_pfn > start_kernel_pfn))
 157                panic("kernel loaded out of ram");
 158
 159        /* Zone start phys-addr must be 2^(MAX_ORDER-1) aligned.
 160           Note that we round this down, not up - node memory
 161           has much larger alignment than 8Mb, so it's safe. */
 162        node_min_pfn &= ~((1UL << (MAX_ORDER-1))-1);
 163
 164        /* We need to know how many physically contiguous pages
 165           we'll need for the bootmap.  */
 166        bootmap_pages = bootmem_bootmap_pages(node_max_pfn-node_min_pfn);
 167
 168        /* Now find a good region where to allocate the bootmap.  */
 169        for_each_mem_cluster(memdesc, cluster, i) {
 170                if (cluster->usage & 3)
 171                        continue;
 172
 173                start = cluster->start_pfn;
 174                end = start + cluster->numpages;
 175
 176                if (start >= node_max_pfn || end <= node_min_pfn)
 177                        continue;
 178
 179                if (end > node_max_pfn)
 180                        end = node_max_pfn;
 181                if (start < node_min_pfn)
 182                        start = node_min_pfn;
 183
 184                if (start < start_kernel_pfn) {
 185                        if (end > end_kernel_pfn
 186                            && end - end_kernel_pfn >= bootmap_pages) {
 187                                bootmap_start = end_kernel_pfn;
 188                                break;
 189                        } else if (end > start_kernel_pfn)
 190                                end = start_kernel_pfn;
 191                } else if (start < end_kernel_pfn)
 192                        start = end_kernel_pfn;
 193                if (end - start >= bootmap_pages) {
 194                        bootmap_start = start;
 195                        break;
 196                }
 197        }
 198
 199        if (bootmap_start == -1)
 200                panic("couldn't find a contiguous place for the bootmap");
 201
 202        /* Allocate the bootmap and mark the whole MM as reserved.  */
 203        bootmap_size = init_bootmem_node(NODE_DATA(nid), bootmap_start,
 204                                         node_min_pfn, node_max_pfn);
 205        DBGDCONT(" bootmap_start %lu, bootmap_size %lu, bootmap_pages %lu\n",
 206                 bootmap_start, bootmap_size, bootmap_pages);
 207
 208        /* Mark the free regions.  */
 209        for_each_mem_cluster(memdesc, cluster, i) {
 210                if (cluster->usage & 3)
 211                        continue;
 212
 213                start = cluster->start_pfn;
 214                end = cluster->start_pfn + cluster->numpages;
 215
 216                if (start >= node_max_pfn || end <= node_min_pfn)
 217                        continue;
 218
 219                if (end > node_max_pfn)
 220                        end = node_max_pfn;
 221                if (start < node_min_pfn)
 222                        start = node_min_pfn;
 223
 224                if (start < start_kernel_pfn) {
 225                        if (end > end_kernel_pfn) {
 226                                free_bootmem_node(NODE_DATA(nid), PFN_PHYS(start),
 227                                             (PFN_PHYS(start_kernel_pfn)
 228                                              - PFN_PHYS(start)));
 229                                printk(" freeing pages %ld:%ld\n",
 230                                       start, start_kernel_pfn);
 231                                start = end_kernel_pfn;
 232                        } else if (end > start_kernel_pfn)
 233                                end = start_kernel_pfn;
 234                } else if (start < end_kernel_pfn)
 235                        start = end_kernel_pfn;
 236                if (start >= end)
 237                        continue;
 238
 239                free_bootmem_node(NODE_DATA(nid), PFN_PHYS(start), PFN_PHYS(end) - PFN_PHYS(start));
 240                printk(" freeing pages %ld:%ld\n", start, end);
 241        }
 242
 243        /* Reserve the bootmap memory.  */
 244        reserve_bootmem_node(NODE_DATA(nid), PFN_PHYS(bootmap_start),
 245                        bootmap_size, BOOTMEM_DEFAULT);
 246        printk(" reserving pages %ld:%ld\n", bootmap_start, bootmap_start+PFN_UP(bootmap_size));
 247
 248        node_set_online(nid);
 249}
 250
 251void __init
 252setup_memory(void *kernel_end)
 253{
 254        int nid;
 255
 256        show_mem_layout();
 257
 258        nodes_clear(node_online_map);
 259
 260        min_low_pfn = ~0UL;
 261        max_low_pfn = 0UL;
 262        for (nid = 0; nid < MAX_NUMNODES; nid++)
 263                setup_memory_node(nid, kernel_end);
 264
 265#ifdef CONFIG_BLK_DEV_INITRD
 266        initrd_start = INITRD_START;
 267        if (initrd_start) {
 268                extern void *move_initrd(unsigned long);
 269
 270                initrd_end = initrd_start+INITRD_SIZE;
 271                printk("Initial ramdisk at: 0x%p (%lu bytes)\n",
 272                       (void *) initrd_start, INITRD_SIZE);
 273
 274                if ((void *)initrd_end > phys_to_virt(PFN_PHYS(max_low_pfn))) {
 275                        if (!move_initrd(PFN_PHYS(max_low_pfn)))
 276                                printk("initrd extends beyond end of memory "
 277                                       "(0x%08lx > 0x%p)\ndisabling initrd\n",
 278                                       initrd_end,
 279                                       phys_to_virt(PFN_PHYS(max_low_pfn)));
 280                } else {
 281                        nid = kvaddr_to_nid(initrd_start);
 282                        reserve_bootmem_node(NODE_DATA(nid),
 283                                             virt_to_phys((void *)initrd_start),
 284                                             INITRD_SIZE, BOOTMEM_DEFAULT);
 285                }
 286        }
 287#endif /* CONFIG_BLK_DEV_INITRD */
 288}
 289
 290void __init paging_init(void)
 291{
 292        unsigned int    nid;
 293        unsigned long   zones_size[MAX_NR_ZONES] = {0, };
 294        unsigned long   dma_local_pfn;
 295
 296        /*
 297         * The old global MAX_DMA_ADDRESS per-arch API doesn't fit
 298         * in the NUMA model, for now we convert it to a pfn and
 299         * we interpret this pfn as a local per-node information.
 300         * This issue isn't very important since none of these machines
 301         * have legacy ISA slots anyways.
 302         */
 303        dma_local_pfn = virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
 304
 305        for_each_online_node(nid) {
 306                bootmem_data_t *bdata = &bootmem_node_data[nid];
 307                unsigned long start_pfn = bdata->node_min_pfn;
 308                unsigned long end_pfn = bdata->node_low_pfn;
 309
 310                if (dma_local_pfn >= end_pfn - start_pfn)
 311                        zones_size[ZONE_DMA] = end_pfn - start_pfn;
 312                else {
 313                        zones_size[ZONE_DMA] = dma_local_pfn;
 314                        zones_size[ZONE_NORMAL] = (end_pfn - start_pfn) - dma_local_pfn;
 315                }
 316                node_set_state(nid, N_NORMAL_MEMORY);
 317                free_area_init_node(nid, zones_size, start_pfn, NULL);
 318        }
 319
 320        /* Initialize the kernel's ZERO_PGE. */
 321        memset((void *)ZERO_PGE, 0, PAGE_SIZE);
 322}
 323