1// SPDX-License-Identifier: GPL-2.0
2#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
3
4#include <linux/kernel.h>
5#include <linux/export.h>
6#include <linux/init.h>
7#include <linux/memblock.h>
8#include <linux/percpu.h>
9#include <linux/kexec.h>
10#include <linux/crash_dump.h>
11#include <linux/smp.h>
12#include <linux/topology.h>
13#include <linux/pfn.h>
14#include <linux/stackprotector.h>
15#include <asm/sections.h>
16#include <asm/processor.h>
17#include <asm/desc.h>
18#include <asm/setup.h>
19#include <asm/mpspec.h>
20#include <asm/apicdef.h>
21#include <asm/highmem.h>
22#include <asm/proto.h>
23#include <asm/cpumask.h>
24#include <asm/cpu.h>
25
26#ifdef CONFIG_X86_64
27#define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
28#else
29#define BOOT_PERCPU_OFFSET 0
30#endif
31
32DEFINE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
33EXPORT_PER_CPU_SYMBOL(this_cpu_off);
34
35unsigned long __per_cpu_offset[NR_CPUS] __ro_after_init = {
36 [0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
37};
38EXPORT_SYMBOL(__per_cpu_offset);
39
40/*
41 * On x86_64 symbols referenced from code should be reachable using
42 * 32bit relocations. Reserve space for static percpu variables in
43 * modules so that they are always served from the first chunk which
44 * is located at the percpu segment base. On x86_32, anything can
45 * address anywhere. No need to reserve space in the first chunk.
46 */
47#ifdef CONFIG_X86_64
48#define PERCPU_FIRST_CHUNK_RESERVE PERCPU_MODULE_RESERVE
49#else
50#define PERCPU_FIRST_CHUNK_RESERVE 0
51#endif
52
53#ifdef CONFIG_X86_32
54/**
55 * pcpu_need_numa - determine percpu allocation needs to consider NUMA
56 *
57 * If NUMA is not configured or there is only one NUMA node available,
58 * there is no reason to consider NUMA. This function determines
59 * whether percpu allocation should consider NUMA or not.
60 *
61 * RETURNS:
62 * true if NUMA should be considered; otherwise, false.
63 */
64static bool __init pcpu_need_numa(void)
65{
66#ifdef CONFIG_NUMA
67 pg_data_t *last = NULL;
68 unsigned int cpu;
69
70 for_each_possible_cpu(cpu) {
71 int node = early_cpu_to_node(cpu);
72
73 if (node_online(node) && NODE_DATA(node) &&
74 last && last != NODE_DATA(node))
75 return true;
76
77 last = NODE_DATA(node);
78 }
79#endif
80 return false;
81}
82#endif
83
84static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
85{
86#ifdef CONFIG_NUMA
87 if (early_cpu_to_node(cpu: from) == early_cpu_to_node(cpu: to))
88 return LOCAL_DISTANCE;
89 else
90 return REMOTE_DISTANCE;
91#else
92 return LOCAL_DISTANCE;
93#endif
94}
95
96static int __init pcpu_cpu_to_node(int cpu)
97{
98 return early_cpu_to_node(cpu);
99}
100
101void __init pcpu_populate_pte(unsigned long addr)
102{
103 populate_extra_pte(vaddr: addr);
104}
105
106static inline void setup_percpu_segment(int cpu)
107{
108#ifdef CONFIG_X86_32
109 struct desc_struct d = GDT_ENTRY_INIT(0x8092, per_cpu_offset(cpu),
110 0xFFFFF);
111
112 write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_PERCPU, &d, DESCTYPE_S);
113#endif
114}
115
116void __init setup_per_cpu_areas(void)
117{
118 unsigned int cpu;
119 unsigned long delta;
120 int rc;
121
122 pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%u nr_node_ids:%u\n",
123 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
124
125 /*
126 * Allocate percpu area. Embedding allocator is our favorite;
127 * however, on NUMA configurations, it can result in very
128 * sparse unit mapping and vmalloc area isn't spacious enough
129 * on 32bit. Use page in that case.
130 */
131#ifdef CONFIG_X86_32
132 if (pcpu_chosen_fc == PCPU_FC_AUTO && pcpu_need_numa())
133 pcpu_chosen_fc = PCPU_FC_PAGE;
134#endif
135 rc = -EINVAL;
136 if (pcpu_chosen_fc != PCPU_FC_PAGE) {
137 const size_t dyn_size = PERCPU_MODULE_RESERVE +
138 PERCPU_DYNAMIC_RESERVE - PERCPU_FIRST_CHUNK_RESERVE;
139 size_t atom_size;
140
141 /*
142 * On 64bit, use PMD_SIZE for atom_size so that embedded
143 * percpu areas are aligned to PMD. This, in the future,
144 * can also allow using PMD mappings in vmalloc area. Use
145 * PAGE_SIZE on 32bit as vmalloc space is highly contended
146 * and large vmalloc area allocs can easily fail.
147 */
148#ifdef CONFIG_X86_64
149 atom_size = PMD_SIZE;
150#else
151 atom_size = PAGE_SIZE;
152#endif
153 rc = pcpu_embed_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
154 dyn_size, atom_size,
155 cpu_distance_fn: pcpu_cpu_distance,
156 cpu_to_nd_fn: pcpu_cpu_to_node);
157 if (rc < 0)
158 pr_warn("%s allocator failed (%d), falling back to page size\n",
159 pcpu_fc_names[pcpu_chosen_fc], rc);
160 }
161 if (rc < 0)
162 rc = pcpu_page_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
163 cpu_to_nd_fn: pcpu_cpu_to_node);
164 if (rc < 0)
165 panic(fmt: "cannot initialize percpu area (err=%d)", rc);
166
167 /* alrighty, percpu areas up and running */
168 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
169 for_each_possible_cpu(cpu) {
170 per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
171 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
172 per_cpu(pcpu_hot.cpu_number, cpu) = cpu;
173 setup_percpu_segment(cpu);
174 /*
175 * Copy data used in early init routines from the
176 * initial arrays to the per cpu data areas. These
177 * arrays then become expendable and the *_early_ptr's
178 * are zeroed indicating that the static arrays are
179 * gone.
180 */
181#ifdef CONFIG_X86_LOCAL_APIC
182 per_cpu(x86_cpu_to_apicid, cpu) =
183 early_per_cpu_map(x86_cpu_to_apicid, cpu);
184 per_cpu(x86_cpu_to_acpiid, cpu) =
185 early_per_cpu_map(x86_cpu_to_acpiid, cpu);
186#endif
187#ifdef CONFIG_NUMA
188 per_cpu(x86_cpu_to_node_map, cpu) =
189 early_per_cpu_map(x86_cpu_to_node_map, cpu);
190 /*
191 * Ensure that the boot cpu numa_node is correct when the boot
192 * cpu is on a node that doesn't have memory installed.
193 * Also cpu_up() will call cpu_to_node() for APs when
194 * MEMORY_HOTPLUG is defined, before per_cpu(numa_node) is set
195 * up later with c_init aka intel_init/amd_init.
196 * So set them all (boot cpu and all APs).
197 */
198 set_cpu_numa_node(cpu, node: early_cpu_to_node(cpu));
199#endif
200 /*
201 * Up to this point, the boot CPU has been using .init.data
202 * area. Reload any changed state for the boot CPU.
203 */
204 if (!cpu)
205 switch_gdt_and_percpu_base(cpu);
206 }
207
208 /* indicate the early static arrays will soon be gone */
209#ifdef CONFIG_X86_LOCAL_APIC
210 early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
211 early_per_cpu_ptr(x86_cpu_to_acpiid) = NULL;
212#endif
213#ifdef CONFIG_NUMA
214 early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
215#endif
216
217 /* Setup node to cpumask map */
218 setup_node_to_cpumask_map();
219
220 /* Setup cpu initialized, callin, callout masks */
221 setup_cpu_local_masks();
222
223 /*
224 * Sync back kernel address range again. We already did this in
225 * setup_arch(), but percpu data also needs to be available in
226 * the smpboot asm and arch_sync_kernel_mappings() doesn't sync to
227 * swapper_pg_dir on 32-bit. The per-cpu mappings need to be available
228 * there too.
229 *
230 * FIXME: Can the later sync in setup_cpu_entry_areas() replace
231 * this call?
232 */
233 sync_initial_page_table();
234}
235

source code of linux/arch/x86/kernel/setup_percpu.c