2 * Copyright (c) 2000, 2003 Silicon Graphics, Inc. All rights reserved.
3 * Copyright (c) 2001 Intel Corp.
4 * Copyright (c) 2001 Tony Luck <tony.luck@intel.com>
5 * Copyright (c) 2002 NEC Corp.
6 * Copyright (c) 2002 Kimio Suganuma <k-suganuma@da.jp.nec.com>
7 * Copyright (c) 2004 Silicon Graphics, Inc
8 * Russ Anderson <rja@sgi.com>
9 * Jesse Barnes <jbarnes@sgi.com>
10 * Jack Steiner <steiner@sgi.com>
14 * Platform initialization for Discontig Memory
17 #include <linux/kernel.h>
19 #include <linux/nmi.h>
20 #include <linux/swap.h>
21 #include <linux/bootmem.h>
22 #include <linux/acpi.h>
23 #include <linux/efi.h>
24 #include <linux/nodemask.h>
25 #include <linux/slab.h>
26 #include <asm/pgalloc.h>
28 #include <asm/meminit.h>
30 #include <asm/sections.h>
33 * Track per-node information needed to setup the boot memory allocator, the
34 * per-node areas, and the real VM.
36 struct early_node_data {
37 struct ia64_node_data *node_data;
38 unsigned long pernode_addr;
39 unsigned long pernode_size;
40 #ifdef CONFIG_ZONE_DMA
41 unsigned long num_dma_physpages;
43 unsigned long min_pfn;
44 unsigned long max_pfn;
47 static struct early_node_data mem_data[MAX_NUMNODES] __initdata;
48 static nodemask_t memory_less_mask __initdata;
50 pg_data_t *pgdat_list[MAX_NUMNODES];
53 * To prevent cache aliasing effects, align per-node structures so that they
54 * start at addresses that are strided by node number.
56 #define MAX_NODE_ALIGN_OFFSET (32 * 1024 * 1024)
57 #define NODEDATA_ALIGN(addr, node) \
58 ((((addr) + 1024*1024-1) & ~(1024*1024-1)) + \
59 (((node)*PERCPU_PAGE_SIZE) & (MAX_NODE_ALIGN_OFFSET - 1)))
62 * build_node_maps - callback to setup bootmem structs for each node
63 * @start: physical start of range
64 * @len: length of range
65 * @node: node where this range resides
67 * We allocate a struct bootmem_data for each piece of memory that we wish to
68 * treat as a virtually contiguous block (i.e. each node). Each such block
69 * must start on an %IA64_GRANULE_SIZE boundary, so we round the address down
70 * if necessary. Any non-existent pages will simply be part of the virtual
71 * memmap. We also update min_low_pfn and max_low_pfn here as we receive
72 * memory ranges from the caller.
74 static int __init build_node_maps(unsigned long start, unsigned long len,
77 unsigned long spfn, epfn, end = start + len;
78 struct bootmem_data *bdp = &bootmem_node_data[node];
80 epfn = GRANULEROUNDUP(end) >> PAGE_SHIFT;
81 spfn = GRANULEROUNDDOWN(start) >> PAGE_SHIFT;
83 if (!bdp->node_low_pfn) {
84 bdp->node_min_pfn = spfn;
85 bdp->node_low_pfn = epfn;
87 bdp->node_min_pfn = min(spfn, bdp->node_min_pfn);
88 bdp->node_low_pfn = max(epfn, bdp->node_low_pfn);
95 * early_nr_cpus_node - return number of cpus on a given node
96 * @node: node to check
98 * Count the number of cpus on @node. We can't use nr_cpus_node() yet because
99 * acpi_boot_init() (which builds the node_to_cpu_mask array) hasn't been
100 * called yet. Note that node 0 will also count all non-existent cpus.
102 static int __meminit early_nr_cpus_node(int node)
106 for_each_possible_early_cpu(cpu)
107 if (node == node_cpuid[cpu].nid)
114 * compute_pernodesize - compute size of pernode data
115 * @node: the node id.
117 static unsigned long __meminit compute_pernodesize(int node)
119 unsigned long pernodesize = 0, cpus;
121 cpus = early_nr_cpus_node(node);
122 pernodesize += PERCPU_PAGE_SIZE * cpus;
123 pernodesize += node * L1_CACHE_BYTES;
124 pernodesize += L1_CACHE_ALIGN(sizeof(pg_data_t));
125 pernodesize += L1_CACHE_ALIGN(sizeof(struct ia64_node_data));
126 pernodesize += L1_CACHE_ALIGN(sizeof(pg_data_t));
127 pernodesize = PAGE_ALIGN(pernodesize);
132 * per_cpu_node_setup - setup per-cpu areas on each node
133 * @cpu_data: per-cpu area on this node
134 * @node: node to setup
136 * Copy the static per-cpu data into the region we just set aside and then
137 * setup __per_cpu_offset for each CPU on this node. Return a pointer to
138 * the end of the area.
140 static void *per_cpu_node_setup(void *cpu_data, int node)
145 for_each_possible_early_cpu(cpu) {
146 void *src = cpu == 0 ? __cpu0_per_cpu : __phys_per_cpu_start;
148 if (node != node_cpuid[cpu].nid)
151 memcpy(__va(cpu_data), src, __per_cpu_end - __per_cpu_start);
152 __per_cpu_offset[cpu] = (char *)__va(cpu_data) -
156 * percpu area for cpu0 is moved from the __init area
157 * which is setup by head.S and used till this point.
158 * Update ar.k3. This move is ensures that percpu
159 * area for cpu0 is on the correct node and its
160 * virtual address isn't insanely far from other
161 * percpu areas which is important for congruent
165 ia64_set_kr(IA64_KR_PER_CPU_DATA,
166 (unsigned long)cpu_data -
167 (unsigned long)__per_cpu_start);
169 cpu_data += PERCPU_PAGE_SIZE;
177 * setup_per_cpu_areas - setup percpu areas
179 * Arch code has already allocated and initialized percpu areas. All
180 * this function has to do is to teach the determined layout to the
181 * dynamic percpu allocator, which happens to be more complex than
182 * creating whole new ones using helpers.
184 void __init setup_per_cpu_areas(void)
186 struct pcpu_alloc_info *ai;
187 struct pcpu_group_info *uninitialized_var(gi);
188 unsigned int *cpu_map;
190 unsigned long base_offset;
192 ssize_t static_size, reserved_size, dyn_size;
193 int node, prev_node, unit, nr_units, rc;
195 ai = pcpu_alloc_alloc_info(MAX_NUMNODES, nr_cpu_ids);
197 panic("failed to allocate pcpu_alloc_info");
198 cpu_map = ai->groups[0].cpu_map;
201 base = (void *)ULONG_MAX;
202 for_each_possible_cpu(cpu)
204 (void *)(__per_cpu_offset[cpu] + __per_cpu_start));
205 base_offset = (void *)__per_cpu_start - base;
207 /* build cpu_map, units are grouped by node */
210 for_each_possible_cpu(cpu)
211 if (node == node_cpuid[cpu].nid)
212 cpu_map[unit++] = cpu;
215 /* set basic parameters */
216 static_size = __per_cpu_end - __per_cpu_start;
217 reserved_size = PERCPU_MODULE_RESERVE;
218 dyn_size = PERCPU_PAGE_SIZE - static_size - reserved_size;
220 panic("percpu area overflow static=%zd reserved=%zd\n",
221 static_size, reserved_size);
223 ai->static_size = static_size;
224 ai->reserved_size = reserved_size;
225 ai->dyn_size = dyn_size;
226 ai->unit_size = PERCPU_PAGE_SIZE;
227 ai->atom_size = PAGE_SIZE;
228 ai->alloc_size = PERCPU_PAGE_SIZE;
231 * CPUs are put into groups according to node. Walk cpu_map
232 * and create new groups at node boundaries.
236 for (unit = 0; unit < nr_units; unit++) {
238 node = node_cpuid[cpu].nid;
240 if (node == prev_node) {
246 gi = &ai->groups[ai->nr_groups++];
248 gi->base_offset = __per_cpu_offset[cpu] + base_offset;
249 gi->cpu_map = &cpu_map[unit];
252 rc = pcpu_setup_first_chunk(ai, base);
254 panic("failed to setup percpu area (err=%d)", rc);
256 pcpu_free_alloc_info(ai);
261 * fill_pernode - initialize pernode data.
262 * @node: the node id.
263 * @pernode: physical address of pernode data
264 * @pernodesize: size of the pernode data
266 static void __init fill_pernode(int node, unsigned long pernode,
267 unsigned long pernodesize)
270 int cpus = early_nr_cpus_node(node);
271 struct bootmem_data *bdp = &bootmem_node_data[node];
273 mem_data[node].pernode_addr = pernode;
274 mem_data[node].pernode_size = pernodesize;
275 memset(__va(pernode), 0, pernodesize);
277 cpu_data = (void *)pernode;
278 pernode += PERCPU_PAGE_SIZE * cpus;
279 pernode += node * L1_CACHE_BYTES;
281 pgdat_list[node] = __va(pernode);
282 pernode += L1_CACHE_ALIGN(sizeof(pg_data_t));
284 mem_data[node].node_data = __va(pernode);
285 pernode += L1_CACHE_ALIGN(sizeof(struct ia64_node_data));
287 pgdat_list[node]->bdata = bdp;
288 pernode += L1_CACHE_ALIGN(sizeof(pg_data_t));
290 cpu_data = per_cpu_node_setup(cpu_data, node);
296 * find_pernode_space - allocate memory for memory map and per-node structures
297 * @start: physical start of range
298 * @len: length of range
299 * @node: node where this range resides
301 * This routine reserves space for the per-cpu data struct, the list of
302 * pg_data_ts and the per-node data struct. Each node will have something like
303 * the following in the first chunk of addr. space large enough to hold it.
305 * ________________________
307 * |~~~~~~~~~~~~~~~~~~~~~~~~| <-- NODEDATA_ALIGN(start, node) for the first
308 * | PERCPU_PAGE_SIZE * | start and length big enough
309 * | cpus_on_this_node | Node 0 will also have entries for all non-existent cpus.
310 * |------------------------|
311 * | local pg_data_t * |
312 * |------------------------|
313 * | local ia64_node_data |
314 * |------------------------|
316 * |________________________|
318 * Once this space has been set aside, the bootmem maps are initialized. We
319 * could probably move the allocation of the per-cpu and ia64_node_data space
320 * outside of this function and use alloc_bootmem_node(), but doing it here
321 * is straightforward and we get the alignments we want so...
323 static int __init find_pernode_space(unsigned long start, unsigned long len,
326 unsigned long spfn, epfn;
327 unsigned long pernodesize = 0, pernode, pages, mapsize;
328 struct bootmem_data *bdp = &bootmem_node_data[node];
330 spfn = start >> PAGE_SHIFT;
331 epfn = (start + len) >> PAGE_SHIFT;
333 pages = bdp->node_low_pfn - bdp->node_min_pfn;
334 mapsize = bootmem_bootmap_pages(pages) << PAGE_SHIFT;
337 * Make sure this memory falls within this node's usable memory
338 * since we may have thrown some away in build_maps().
340 if (spfn < bdp->node_min_pfn || epfn > bdp->node_low_pfn)
343 /* Don't setup this node's local space twice... */
344 if (mem_data[node].pernode_addr)
348 * Calculate total size needed, incl. what's necessary
349 * for good alignment and alias prevention.
351 pernodesize = compute_pernodesize(node);
352 pernode = NODEDATA_ALIGN(start, node);
354 /* Is this range big enough for what we want to store here? */
355 if (start + len > (pernode + pernodesize + mapsize))
356 fill_pernode(node, pernode, pernodesize);
362 * free_node_bootmem - free bootmem allocator memory for use
363 * @start: physical start of range
364 * @len: length of range
365 * @node: node where this range resides
367 * Simply calls the bootmem allocator to free the specified ranged from
368 * the given pg_data_t's bdata struct. After this function has been called
369 * for all the entries in the EFI memory map, the bootmem allocator will
370 * be ready to service allocation requests.
372 static int __init free_node_bootmem(unsigned long start, unsigned long len,
375 free_bootmem_node(pgdat_list[node], start, len);
381 * reserve_pernode_space - reserve memory for per-node space
383 * Reserve the space used by the bootmem maps & per-node space in the boot
384 * allocator so that when we actually create the real mem maps we don't
387 static void __init reserve_pernode_space(void)
389 unsigned long base, size, pages;
390 struct bootmem_data *bdp;
393 for_each_online_node(node) {
394 pg_data_t *pdp = pgdat_list[node];
396 if (node_isset(node, memory_less_mask))
401 /* First the bootmem_map itself */
402 pages = bdp->node_low_pfn - bdp->node_min_pfn;
403 size = bootmem_bootmap_pages(pages) << PAGE_SHIFT;
404 base = __pa(bdp->node_bootmem_map);
405 reserve_bootmem_node(pdp, base, size, BOOTMEM_DEFAULT);
407 /* Now the per-node space */
408 size = mem_data[node].pernode_size;
409 base = __pa(mem_data[node].pernode_addr);
410 reserve_bootmem_node(pdp, base, size, BOOTMEM_DEFAULT);
414 static void __meminit scatter_node_data(void)
420 * for_each_online_node() can't be used at here.
421 * node_online_map is not set for hot-added nodes at this time,
422 * because we are halfway through initialization of the new node's
423 * structures. If for_each_online_node() is used, a new node's
424 * pg_data_ptrs will be not initialized. Instead of using it,
425 * pgdat_list[] is checked.
427 for_each_node(node) {
428 if (pgdat_list[node]) {
429 dst = LOCAL_DATA_ADDR(pgdat_list[node])->pg_data_ptrs;
430 memcpy(dst, pgdat_list, sizeof(pgdat_list));
436 * initialize_pernode_data - fixup per-cpu & per-node pointers
438 * Each node's per-node area has a copy of the global pg_data_t list, so
439 * we copy that to each node here, as well as setting the per-cpu pointer
440 * to the local node data structure. The active_cpus field of the per-node
441 * structure gets setup by the platform_cpu_init() function later.
443 static void __init initialize_pernode_data(void)
450 /* Set the node_data pointer for each per-cpu struct */
451 for_each_possible_early_cpu(cpu) {
452 node = node_cpuid[cpu].nid;
453 per_cpu(ia64_cpu_info, cpu).node_data =
454 mem_data[node].node_data;
458 struct cpuinfo_ia64 *cpu0_cpu_info;
460 node = node_cpuid[cpu].nid;
461 cpu0_cpu_info = (struct cpuinfo_ia64 *)(__phys_per_cpu_start +
462 ((char *)&ia64_cpu_info - __per_cpu_start));
463 cpu0_cpu_info->node_data = mem_data[node].node_data;
465 #endif /* CONFIG_SMP */
469 * memory_less_node_alloc - * attempt to allocate memory on the best NUMA slit
470 * node but fall back to any other node when __alloc_bootmem_node fails
473 * @pernodesize: size of this node's pernode data
475 static void __init *memory_less_node_alloc(int nid, unsigned long pernodesize)
479 int bestnode = -1, node, anynode = 0;
481 for_each_online_node(node) {
482 if (node_isset(node, memory_less_mask))
484 else if (node_distance(nid, node) < best) {
485 best = node_distance(nid, node);
494 ptr = __alloc_bootmem_node(pgdat_list[bestnode], pernodesize,
495 PERCPU_PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
501 * memory_less_nodes - allocate and initialize CPU only nodes pernode
504 static void __init memory_less_nodes(void)
506 unsigned long pernodesize;
510 for_each_node_mask(node, memory_less_mask) {
511 pernodesize = compute_pernodesize(node);
512 pernode = memory_less_node_alloc(node, pernodesize);
513 fill_pernode(node, __pa(pernode), pernodesize);
520 * find_memory - walk the EFI memory map and setup the bootmem allocator
522 * Called early in boot to setup the bootmem allocator, and to
523 * allocate the per-cpu and per-node structures.
525 void __init find_memory(void)
531 if (num_online_nodes() == 0) {
532 printk(KERN_ERR "node info missing!\n");
536 nodes_or(memory_less_mask, memory_less_mask, node_online_map);
540 /* These actually end up getting called by call_pernode_memory() */
541 efi_memmap_walk(filter_rsvd_memory, build_node_maps);
542 efi_memmap_walk(filter_rsvd_memory, find_pernode_space);
543 efi_memmap_walk(find_max_min_low_pfn, NULL);
545 for_each_online_node(node)
546 if (bootmem_node_data[node].node_low_pfn) {
547 node_clear(node, memory_less_mask);
548 mem_data[node].min_pfn = ~0UL;
551 efi_memmap_walk(filter_memory, register_active_ranges);
554 * Initialize the boot memory maps in reverse order since that's
555 * what the bootmem allocator expects
557 for (node = MAX_NUMNODES - 1; node >= 0; node--) {
558 unsigned long pernode, pernodesize, map;
559 struct bootmem_data *bdp;
561 if (!node_online(node))
563 else if (node_isset(node, memory_less_mask))
566 bdp = &bootmem_node_data[node];
567 pernode = mem_data[node].pernode_addr;
568 pernodesize = mem_data[node].pernode_size;
569 map = pernode + pernodesize;
571 init_bootmem_node(pgdat_list[node],
577 efi_memmap_walk(filter_rsvd_memory, free_node_bootmem);
579 reserve_pernode_space();
581 initialize_pernode_data();
583 max_pfn = max_low_pfn;
590 * per_cpu_init - setup per-cpu variables
592 * find_pernode_space() does most of this already, we just need to set
593 * local_per_cpu_offset
595 void *per_cpu_init(void)
598 static int first_time = 1;
602 for_each_possible_early_cpu(cpu)
603 per_cpu(local_per_cpu_offset, cpu) = __per_cpu_offset[cpu];
606 return __per_cpu_start + __per_cpu_offset[smp_processor_id()];
608 #endif /* CONFIG_SMP */
611 * show_mem - give short summary of memory stats
613 * Shows a simple page count of reserved and used pages in the system.
614 * For discontig machines, it does this on a per-pgdat basis.
616 void show_mem(unsigned int filter)
618 int i, total_reserved = 0;
619 int total_shared = 0, total_cached = 0;
620 unsigned long total_present = 0;
623 printk(KERN_INFO "Mem-info:\n");
624 show_free_areas(filter);
625 if (filter & SHOW_MEM_FILTER_PAGE_COUNT)
627 printk(KERN_INFO "Node memory in pages:\n");
628 for_each_online_pgdat(pgdat) {
629 unsigned long present;
631 int shared = 0, cached = 0, reserved = 0;
632 int nid = pgdat->node_id;
634 if (skip_free_areas_node(filter, nid))
636 pgdat_resize_lock(pgdat, &flags);
637 present = pgdat->node_present_pages;
638 for(i = 0; i < pgdat->node_spanned_pages; i++) {
640 if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
641 touch_nmi_watchdog();
642 if (pfn_valid(pgdat->node_start_pfn + i))
643 page = pfn_to_page(pgdat->node_start_pfn + i);
645 i = vmemmap_find_next_valid_pfn(nid, i) - 1;
648 if (PageReserved(page))
650 else if (PageSwapCache(page))
652 else if (page_count(page))
653 shared += page_count(page)-1;
655 pgdat_resize_unlock(pgdat, &flags);
656 total_present += present;
657 total_reserved += reserved;
658 total_cached += cached;
659 total_shared += shared;
660 printk(KERN_INFO "Node %4d: RAM: %11ld, rsvd: %8d, "
661 "shrd: %10d, swpd: %10d\n", nid,
662 present, reserved, shared, cached);
664 printk(KERN_INFO "%ld pages of RAM\n", total_present);
665 printk(KERN_INFO "%d reserved pages\n", total_reserved);
666 printk(KERN_INFO "%d pages shared\n", total_shared);
667 printk(KERN_INFO "%d pages swap cached\n", total_cached);
668 printk(KERN_INFO "Total of %ld pages in page table cache\n",
669 quicklist_total_size());
670 printk(KERN_INFO "%ld free buffer pages\n", nr_free_buffer_pages());
674 * call_pernode_memory - use SRAT to call callback functions with node info
675 * @start: physical start of range
676 * @len: length of range
677 * @arg: function to call for each range
679 * efi_memmap_walk() knows nothing about layout of memory across nodes. Find
680 * out to which node a block of memory belongs. Ignore memory that we cannot
681 * identify, and split blocks that run across multiple nodes.
683 * Take this opportunity to round the start address up and the end address
684 * down to page boundaries.
686 void call_pernode_memory(unsigned long start, unsigned long len, void *arg)
688 unsigned long rs, re, end = start + len;
689 void (*func)(unsigned long, unsigned long, int);
692 start = PAGE_ALIGN(start);
699 if (!num_node_memblks) {
700 /* No SRAT table, so assume one node (node 0) */
702 (*func)(start, end - start, 0);
706 for (i = 0; i < num_node_memblks; i++) {
707 rs = max(start, node_memblk[i].start_paddr);
708 re = min(end, node_memblk[i].start_paddr +
709 node_memblk[i].size);
712 (*func)(rs, re - rs, node_memblk[i].nid);
720 * count_node_pages - callback to build per-node memory info structures
721 * @start: physical start of range
722 * @len: length of range
723 * @node: node where this range resides
725 * Each node has it's own number of physical pages, DMAable pages, start, and
726 * end page frame number. This routine will be called by call_pernode_memory()
727 * for each piece of usable memory and will setup these values for each node.
728 * Very similar to build_maps().
730 static __init int count_node_pages(unsigned long start, unsigned long len, int node)
732 unsigned long end = start + len;
734 #ifdef CONFIG_ZONE_DMA
735 if (start <= __pa(MAX_DMA_ADDRESS))
736 mem_data[node].num_dma_physpages +=
737 (min(end, __pa(MAX_DMA_ADDRESS)) - start) >>PAGE_SHIFT;
739 start = GRANULEROUNDDOWN(start);
740 end = GRANULEROUNDUP(end);
741 mem_data[node].max_pfn = max(mem_data[node].max_pfn,
743 mem_data[node].min_pfn = min(mem_data[node].min_pfn,
744 start >> PAGE_SHIFT);
750 * paging_init - setup page tables
752 * paging_init() sets up the page tables for each node of the system and frees
753 * the bootmem allocator memory for general use.
755 void __init paging_init(void)
757 unsigned long max_dma;
758 unsigned long pfn_offset = 0;
759 unsigned long max_pfn = 0;
761 unsigned long max_zone_pfns[MAX_NR_ZONES];
763 max_dma = virt_to_phys((void *) MAX_DMA_ADDRESS) >> PAGE_SHIFT;
765 efi_memmap_walk(filter_rsvd_memory, count_node_pages);
767 sparse_memory_present_with_active_regions(MAX_NUMNODES);
770 #ifdef CONFIG_VIRTUAL_MEM_MAP
771 VMALLOC_END -= PAGE_ALIGN(ALIGN(max_low_pfn, MAX_ORDER_NR_PAGES) *
772 sizeof(struct page));
773 vmem_map = (struct page *) VMALLOC_END;
774 efi_memmap_walk(create_mem_map_page_table, NULL);
775 printk("Virtual mem_map starts at 0x%p\n", vmem_map);
778 for_each_online_node(node) {
779 pfn_offset = mem_data[node].min_pfn;
781 #ifdef CONFIG_VIRTUAL_MEM_MAP
782 NODE_DATA(node)->node_mem_map = vmem_map + pfn_offset;
784 if (mem_data[node].max_pfn > max_pfn)
785 max_pfn = mem_data[node].max_pfn;
788 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
789 #ifdef CONFIG_ZONE_DMA
790 max_zone_pfns[ZONE_DMA] = max_dma;
792 max_zone_pfns[ZONE_NORMAL] = max_pfn;
793 free_area_init_nodes(max_zone_pfns);
795 zero_page_memmap_ptr = virt_to_page(ia64_imva(empty_zero_page));
798 #ifdef CONFIG_MEMORY_HOTPLUG
799 pg_data_t *arch_alloc_nodedata(int nid)
801 unsigned long size = compute_pernodesize(nid);
803 return kzalloc(size, GFP_KERNEL);
806 void arch_free_nodedata(pg_data_t *pgdat)
811 void arch_refresh_nodedata(int update_node, pg_data_t *update_pgdat)
813 pgdat_list[update_node] = update_pgdat;
818 #ifdef CONFIG_SPARSEMEM_VMEMMAP
819 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
821 return vmemmap_populate_basepages(start, end, node);
824 void vmemmap_free(unsigned long start, unsigned long end)