xref: /linux/arch/x86/kernel/setup_percpu.c (revision ea8a163e02d6925773129e2dd86e419e491b791d)
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 <asm/sections.h>
15 #include <asm/processor.h>
16 #include <asm/desc.h>
17 #include <asm/setup.h>
18 #include <asm/mpspec.h>
19 #include <asm/apicdef.h>
20 #include <asm/highmem.h>
21 #include <asm/proto.h>
22 #include <asm/cpumask.h>
23 #include <asm/cpu.h>
24 #include <asm/stackprotector.h>
25 
26 DEFINE_PER_CPU_READ_MOSTLY(int, cpu_number);
27 EXPORT_PER_CPU_SYMBOL(cpu_number);
28 
29 #ifdef CONFIG_X86_64
30 #define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
31 #else
32 #define BOOT_PERCPU_OFFSET 0
33 #endif
34 
35 DEFINE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
36 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
37 
38 unsigned long __per_cpu_offset[NR_CPUS] __ro_after_init = {
39 	[0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
40 };
41 EXPORT_SYMBOL(__per_cpu_offset);
42 
43 /*
44  * On x86_64 symbols referenced from code should be reachable using
45  * 32bit relocations.  Reserve space for static percpu variables in
46  * modules so that they are always served from the first chunk which
47  * is located at the percpu segment base.  On x86_32, anything can
48  * address anywhere.  No need to reserve space in the first chunk.
49  */
50 #ifdef CONFIG_X86_64
51 #define PERCPU_FIRST_CHUNK_RESERVE	PERCPU_MODULE_RESERVE
52 #else
53 #define PERCPU_FIRST_CHUNK_RESERVE	0
54 #endif
55 
56 #ifdef CONFIG_X86_32
57 /**
58  * pcpu_need_numa - determine percpu allocation needs to consider NUMA
59  *
60  * If NUMA is not configured or there is only one NUMA node available,
61  * there is no reason to consider NUMA.  This function determines
62  * whether percpu allocation should consider NUMA or not.
63  *
64  * RETURNS:
65  * true if NUMA should be considered; otherwise, false.
66  */
67 static bool __init pcpu_need_numa(void)
68 {
69 #ifdef CONFIG_NUMA
70 	pg_data_t *last = NULL;
71 	unsigned int cpu;
72 
73 	for_each_possible_cpu(cpu) {
74 		int node = early_cpu_to_node(cpu);
75 
76 		if (node_online(node) && NODE_DATA(node) &&
77 		    last && last != NODE_DATA(node))
78 			return true;
79 
80 		last = NODE_DATA(node);
81 	}
82 #endif
83 	return false;
84 }
85 #endif
86 
87 /**
88  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
89  * @cpu: cpu to allocate for
90  * @size: size allocation in bytes
91  * @align: alignment
92  *
93  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
94  * does the right thing for NUMA regardless of the current
95  * configuration.
96  *
97  * RETURNS:
98  * Pointer to the allocated area on success, NULL on failure.
99  */
100 static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size,
101 					unsigned long align)
102 {
103 	const unsigned long goal = __pa(MAX_DMA_ADDRESS);
104 #ifdef CONFIG_NUMA
105 	int node = early_cpu_to_node(cpu);
106 	void *ptr;
107 
108 	if (!node_online(node) || !NODE_DATA(node)) {
109 		ptr = memblock_alloc_from(size, align, goal);
110 		pr_info("cpu %d has no node %d or node-local memory\n",
111 			cpu, node);
112 		pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
113 			 cpu, size, __pa(ptr));
114 	} else {
115 		ptr = memblock_alloc_try_nid(size, align, goal,
116 					     MEMBLOCK_ALLOC_ACCESSIBLE,
117 					     node);
118 
119 		pr_debug("per cpu data for cpu%d %lu bytes on node%d at %016lx\n",
120 			 cpu, size, node, __pa(ptr));
121 	}
122 	return ptr;
123 #else
124 	return memblock_alloc_from(size, align, goal);
125 #endif
126 }
127 
128 /*
129  * Helpers for first chunk memory allocation
130  */
131 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
132 {
133 	return pcpu_alloc_bootmem(cpu, size, align);
134 }
135 
136 static void __init pcpu_fc_free(void *ptr, size_t size)
137 {
138 	memblock_free(ptr, size);
139 }
140 
141 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
142 {
143 #ifdef CONFIG_NUMA
144 	if (early_cpu_to_node(from) == early_cpu_to_node(to))
145 		return LOCAL_DISTANCE;
146 	else
147 		return REMOTE_DISTANCE;
148 #else
149 	return LOCAL_DISTANCE;
150 #endif
151 }
152 
153 static void __init pcpup_populate_pte(unsigned long addr)
154 {
155 	populate_extra_pte(addr);
156 }
157 
158 static inline void setup_percpu_segment(int cpu)
159 {
160 #ifdef CONFIG_X86_32
161 	struct desc_struct d = GDT_ENTRY_INIT(0x8092, per_cpu_offset(cpu),
162 					      0xFFFFF);
163 
164 	write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_PERCPU, &d, DESCTYPE_S);
165 #endif
166 }
167 
168 void __init setup_per_cpu_areas(void)
169 {
170 	unsigned int cpu;
171 	unsigned long delta;
172 	int rc;
173 
174 	pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%u nr_node_ids:%u\n",
175 		NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
176 
177 	/*
178 	 * Allocate percpu area.  Embedding allocator is our favorite;
179 	 * however, on NUMA configurations, it can result in very
180 	 * sparse unit mapping and vmalloc area isn't spacious enough
181 	 * on 32bit.  Use page in that case.
182 	 */
183 #ifdef CONFIG_X86_32
184 	if (pcpu_chosen_fc == PCPU_FC_AUTO && pcpu_need_numa())
185 		pcpu_chosen_fc = PCPU_FC_PAGE;
186 #endif
187 	rc = -EINVAL;
188 	if (pcpu_chosen_fc != PCPU_FC_PAGE) {
189 		const size_t dyn_size = PERCPU_MODULE_RESERVE +
190 			PERCPU_DYNAMIC_RESERVE - PERCPU_FIRST_CHUNK_RESERVE;
191 		size_t atom_size;
192 
193 		/*
194 		 * On 64bit, use PMD_SIZE for atom_size so that embedded
195 		 * percpu areas are aligned to PMD.  This, in the future,
196 		 * can also allow using PMD mappings in vmalloc area.  Use
197 		 * PAGE_SIZE on 32bit as vmalloc space is highly contended
198 		 * and large vmalloc area allocs can easily fail.
199 		 */
200 #ifdef CONFIG_X86_64
201 		atom_size = PMD_SIZE;
202 #else
203 		atom_size = PAGE_SIZE;
204 #endif
205 		rc = pcpu_embed_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
206 					    dyn_size, atom_size,
207 					    pcpu_cpu_distance,
208 					    pcpu_fc_alloc, pcpu_fc_free);
209 		if (rc < 0)
210 			pr_warn("%s allocator failed (%d), falling back to page size\n",
211 				pcpu_fc_names[pcpu_chosen_fc], rc);
212 	}
213 	if (rc < 0)
214 		rc = pcpu_page_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
215 					   pcpu_fc_alloc, pcpu_fc_free,
216 					   pcpup_populate_pte);
217 	if (rc < 0)
218 		panic("cannot initialize percpu area (err=%d)", rc);
219 
220 	/* alrighty, percpu areas up and running */
221 	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
222 	for_each_possible_cpu(cpu) {
223 		per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
224 		per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
225 		per_cpu(cpu_number, cpu) = cpu;
226 		setup_percpu_segment(cpu);
227 		/*
228 		 * Copy data used in early init routines from the
229 		 * initial arrays to the per cpu data areas.  These
230 		 * arrays then become expendable and the *_early_ptr's
231 		 * are zeroed indicating that the static arrays are
232 		 * gone.
233 		 */
234 #ifdef CONFIG_X86_LOCAL_APIC
235 		per_cpu(x86_cpu_to_apicid, cpu) =
236 			early_per_cpu_map(x86_cpu_to_apicid, cpu);
237 		per_cpu(x86_bios_cpu_apicid, cpu) =
238 			early_per_cpu_map(x86_bios_cpu_apicid, cpu);
239 		per_cpu(x86_cpu_to_acpiid, cpu) =
240 			early_per_cpu_map(x86_cpu_to_acpiid, cpu);
241 #endif
242 #ifdef CONFIG_X86_32
243 		per_cpu(x86_cpu_to_logical_apicid, cpu) =
244 			early_per_cpu_map(x86_cpu_to_logical_apicid, cpu);
245 #endif
246 #ifdef CONFIG_NUMA
247 		per_cpu(x86_cpu_to_node_map, cpu) =
248 			early_per_cpu_map(x86_cpu_to_node_map, cpu);
249 		/*
250 		 * Ensure that the boot cpu numa_node is correct when the boot
251 		 * cpu is on a node that doesn't have memory installed.
252 		 * Also cpu_up() will call cpu_to_node() for APs when
253 		 * MEMORY_HOTPLUG is defined, before per_cpu(numa_node) is set
254 		 * up later with c_init aka intel_init/amd_init.
255 		 * So set them all (boot cpu and all APs).
256 		 */
257 		set_cpu_numa_node(cpu, early_cpu_to_node(cpu));
258 #endif
259 		/*
260 		 * Up to this point, the boot CPU has been using .init.data
261 		 * area.  Reload any changed state for the boot CPU.
262 		 */
263 		if (!cpu)
264 			switch_to_new_gdt(cpu);
265 	}
266 
267 	/* indicate the early static arrays will soon be gone */
268 #ifdef CONFIG_X86_LOCAL_APIC
269 	early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
270 	early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
271 	early_per_cpu_ptr(x86_cpu_to_acpiid) = NULL;
272 #endif
273 #ifdef CONFIG_X86_32
274 	early_per_cpu_ptr(x86_cpu_to_logical_apicid) = NULL;
275 #endif
276 #ifdef CONFIG_NUMA
277 	early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
278 #endif
279 
280 	/* Setup node to cpumask map */
281 	setup_node_to_cpumask_map();
282 
283 	/* Setup cpu initialized, callin, callout masks */
284 	setup_cpu_local_masks();
285 
286 	/*
287 	 * Sync back kernel address range again.  We already did this in
288 	 * setup_arch(), but percpu data also needs to be available in
289 	 * the smpboot asm and arch_sync_kernel_mappings() doesn't sync to
290 	 * swapper_pg_dir on 32-bit. The per-cpu mappings need to be available
291 	 * there too.
292 	 *
293 	 * FIXME: Can the later sync in setup_cpu_entry_areas() replace
294 	 * this call?
295 	 */
296 	sync_initial_page_table();
297 }
298