1 #include <linux/kernel.h>
2 #include <linux/module.h>
3 #include <linux/init.h>
4 #include <linux/bootmem.h>
5 #include <linux/percpu.h>
6 #include <linux/kexec.h>
7 #include <linux/crash_dump.h>
9 #include <linux/topology.h>
10 #include <linux/pfn.h>
11 #include <asm/sections.h>
12 #include <asm/processor.h>
13 #include <asm/setup.h>
14 #include <asm/mpspec.h>
15 #include <asm/apicdef.h>
16 #include <asm/highmem.h>
17 #include <asm/proto.h>
18 #include <asm/cpumask.h>
20 #include <asm/stackprotector.h>
22 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
23 # define DBG(x...) printk(KERN_DEBUG x)
28 DEFINE_PER_CPU(int, cpu_number);
29 EXPORT_PER_CPU_SYMBOL(cpu_number);
32 #define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
34 #define BOOT_PERCPU_OFFSET 0
37 DEFINE_PER_CPU(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
38 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
40 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly = {
41 [0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
43 EXPORT_SYMBOL(__per_cpu_offset);
46 * On x86_64 symbols referenced from code should be reachable using
47 * 32bit relocations. Reserve space for static percpu variables in
48 * modules so that they are always served from the first chunk which
49 * is located at the percpu segment base. On x86_32, anything can
50 * address anywhere. No need to reserve space in the first chunk.
53 #define PERCPU_FIRST_CHUNK_RESERVE PERCPU_MODULE_RESERVE
55 #define PERCPU_FIRST_CHUNK_RESERVE 0
59 * pcpu_need_numa - determine percpu allocation needs to consider NUMA
61 * If NUMA is not configured or there is only one NUMA node available,
62 * there is no reason to consider NUMA. This function determines
63 * whether percpu allocation should consider NUMA or not.
66 * true if NUMA should be considered; otherwise, false.
68 static bool __init pcpu_need_numa(void)
70 #ifdef CONFIG_NEED_MULTIPLE_NODES
71 pg_data_t *last = NULL;
74 for_each_possible_cpu(cpu) {
75 int node = early_cpu_to_node(cpu);
77 if (node_online(node) && NODE_DATA(node) &&
78 last && last != NODE_DATA(node))
81 last = NODE_DATA(node);
88 * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
89 * @cpu: cpu to allocate for
90 * @size: size allocation in bytes
93 * Allocate @size bytes aligned at @align for cpu @cpu. This wrapper
94 * does the right thing for NUMA regardless of the current
98 * Pointer to the allocated area on success, NULL on failure.
100 static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size,
103 const unsigned long goal = __pa(MAX_DMA_ADDRESS);
104 #ifdef CONFIG_NEED_MULTIPLE_NODES
105 int node = early_cpu_to_node(cpu);
108 if (!node_online(node) || !NODE_DATA(node)) {
109 ptr = __alloc_bootmem_nopanic(size, align, goal);
110 pr_info("cpu %d has no node %d or node-local memory\n",
112 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
113 cpu, size, __pa(ptr));
115 ptr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
117 pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
118 "%016lx\n", cpu, size, node, __pa(ptr));
122 return __alloc_bootmem_nopanic(size, align, goal);
127 * Helpers for first chunk memory allocation
129 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size)
131 return pcpu_alloc_bootmem(cpu, size, size);
134 static void __init pcpu_fc_free(void *ptr, size_t size)
136 free_bootmem(__pa(ptr), size);
140 * Large page remapping allocator
142 #ifdef CONFIG_NEED_MULTIPLE_NODES
143 static void __init pcpul_map(void *ptr, size_t size, void *addr)
147 pmd = populate_extra_pmd((unsigned long)addr);
148 pmd_v = pfn_pmd(page_to_pfn(virt_to_page(ptr)), PAGE_KERNEL_LARGE);
152 static ssize_t __init setup_pcpu_lpage(size_t static_size, bool chosen)
154 size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
157 size_t vm_size = VMALLOC_END - VMALLOC_START;
158 size_t tot_size = num_possible_cpus() * PMD_SIZE;
160 /* on non-NUMA, embedding is better */
161 if (!pcpu_need_numa())
164 /* don't consume more than 20% of vmalloc area */
165 if (tot_size > vm_size / 5) {
166 pr_info("PERCPU: too large chunk size %zuMB for "
167 "large page remap\n", tot_size >> 20);
174 pr_warning("PERCPU: lpage allocator requires PSE\n");
178 return pcpu_lpage_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE,
179 reserve - PERCPU_FIRST_CHUNK_RESERVE,
181 pcpu_fc_alloc, pcpu_fc_free, pcpul_map);
184 static ssize_t __init setup_pcpu_lpage(size_t static_size, bool chosen)
191 * Embedding allocator
193 * The first chunk is sized to just contain the static area plus
194 * module and dynamic reserves and embedded into linear physical
195 * mapping so that it can use PMD mapping without additional TLB
198 static ssize_t __init setup_pcpu_embed(size_t static_size, bool chosen)
200 size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
203 * If large page isn't supported, there's no benefit in doing
204 * this. Also, embedding allocation doesn't play well with
207 if (!chosen && (!cpu_has_pse || pcpu_need_numa()))
210 return pcpu_embed_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE,
211 reserve - PERCPU_FIRST_CHUNK_RESERVE);
217 * Boring fallback 4k allocator. This allocator puts more pressure on
218 * PTE TLBs but other than that behaves nicely on both UMA and NUMA.
220 static void __init pcpu4k_populate_pte(unsigned long addr)
222 populate_extra_pte(addr);
225 static ssize_t __init setup_pcpu_4k(size_t static_size)
227 return pcpu_4k_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE,
228 pcpu_fc_alloc, pcpu_fc_free,
229 pcpu4k_populate_pte);
232 /* for explicit first chunk allocator selection */
233 static char pcpu_chosen_alloc[16] __initdata;
235 static int __init percpu_alloc_setup(char *str)
237 strncpy(pcpu_chosen_alloc, str, sizeof(pcpu_chosen_alloc) - 1);
240 early_param("percpu_alloc", percpu_alloc_setup);
242 static inline void setup_percpu_segment(int cpu)
245 struct desc_struct gdt;
247 pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
248 0x2 | DESCTYPE_S, 0x8);
250 write_gdt_entry(get_cpu_gdt_table(cpu),
251 GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S);
255 void __init setup_per_cpu_areas(void)
257 size_t static_size = __per_cpu_end - __per_cpu_start;
260 size_t pcpu_unit_size;
263 pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
264 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
267 * Allocate percpu area. If PSE is supported, try to make use
268 * of large page mappings. Please read comments on top of
269 * each allocator for details.
272 if (strlen(pcpu_chosen_alloc)) {
273 if (strcmp(pcpu_chosen_alloc, "4k")) {
274 if (!strcmp(pcpu_chosen_alloc, "lpage"))
275 ret = setup_pcpu_lpage(static_size, true);
276 else if (!strcmp(pcpu_chosen_alloc, "embed"))
277 ret = setup_pcpu_embed(static_size, true);
279 pr_warning("PERCPU: unknown allocator %s "
280 "specified\n", pcpu_chosen_alloc);
282 pr_warning("PERCPU: %s allocator failed (%zd), "
283 "falling back to 4k\n",
284 pcpu_chosen_alloc, ret);
287 ret = setup_pcpu_lpage(static_size, false);
289 ret = setup_pcpu_embed(static_size, false);
292 ret = setup_pcpu_4k(static_size);
294 panic("cannot allocate static percpu area (%zu bytes, err=%zd)",
297 pcpu_unit_size = ret;
299 /* alrighty, percpu areas up and running */
300 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
301 for_each_possible_cpu(cpu) {
302 per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size;
303 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
304 per_cpu(cpu_number, cpu) = cpu;
305 setup_percpu_segment(cpu);
306 setup_stack_canary_segment(cpu);
308 * Copy data used in early init routines from the
309 * initial arrays to the per cpu data areas. These
310 * arrays then become expendable and the *_early_ptr's
311 * are zeroed indicating that the static arrays are
314 #ifdef CONFIG_X86_LOCAL_APIC
315 per_cpu(x86_cpu_to_apicid, cpu) =
316 early_per_cpu_map(x86_cpu_to_apicid, cpu);
317 per_cpu(x86_bios_cpu_apicid, cpu) =
318 early_per_cpu_map(x86_bios_cpu_apicid, cpu);
321 per_cpu(irq_stack_ptr, cpu) =
322 per_cpu(irq_stack_union.irq_stack, cpu) +
325 per_cpu(x86_cpu_to_node_map, cpu) =
326 early_per_cpu_map(x86_cpu_to_node_map, cpu);
330 * Up to this point, the boot CPU has been using .data.init
331 * area. Reload any changed state for the boot CPU.
333 if (cpu == boot_cpu_id)
334 switch_to_new_gdt(cpu);
337 /* indicate the early static arrays will soon be gone */
338 #ifdef CONFIG_X86_LOCAL_APIC
339 early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
340 early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
342 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
343 early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
346 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
348 * make sure boot cpu node_number is right, when boot cpu is on the
349 * node that doesn't have mem installed
351 per_cpu(node_number, boot_cpu_id) = cpu_to_node(boot_cpu_id);
354 /* Setup node to cpumask map */
355 setup_node_to_cpumask_map();
357 /* Setup cpu initialized, callin, callout masks */
358 setup_cpu_local_masks();