xref: /openbmc/linux/arch/x86/kernel/setup_percpu.c (revision 1c2dd16a)
1 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
2 
3 #include <linux/kernel.h>
4 #include <linux/export.h>
5 #include <linux/init.h>
6 #include <linux/bootmem.h>
7 #include <linux/percpu.h>
8 #include <linux/kexec.h>
9 #include <linux/crash_dump.h>
10 #include <linux/smp.h>
11 #include <linux/topology.h>
12 #include <linux/pfn.h>
13 #include <asm/sections.h>
14 #include <asm/processor.h>
15 #include <asm/desc.h>
16 #include <asm/setup.h>
17 #include <asm/mpspec.h>
18 #include <asm/apicdef.h>
19 #include <asm/highmem.h>
20 #include <asm/proto.h>
21 #include <asm/cpumask.h>
22 #include <asm/cpu.h>
23 #include <asm/stackprotector.h>
24 
25 DEFINE_PER_CPU_READ_MOSTLY(int, cpu_number);
26 EXPORT_PER_CPU_SYMBOL(cpu_number);
27 
28 #ifdef CONFIG_X86_64
29 #define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
30 #else
31 #define BOOT_PERCPU_OFFSET 0
32 #endif
33 
34 DEFINE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
35 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
36 
37 unsigned long __per_cpu_offset[NR_CPUS] __ro_after_init = {
38 	[0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
39 };
40 EXPORT_SYMBOL(__per_cpu_offset);
41 
42 /*
43  * On x86_64 symbols referenced from code should be reachable using
44  * 32bit relocations.  Reserve space for static percpu variables in
45  * modules so that they are always served from the first chunk which
46  * is located at the percpu segment base.  On x86_32, anything can
47  * address anywhere.  No need to reserve space in the first chunk.
48  */
49 #ifdef CONFIG_X86_64
50 #define PERCPU_FIRST_CHUNK_RESERVE	PERCPU_MODULE_RESERVE
51 #else
52 #define PERCPU_FIRST_CHUNK_RESERVE	0
53 #endif
54 
55 #ifdef CONFIG_X86_32
56 /**
57  * pcpu_need_numa - determine percpu allocation needs to consider NUMA
58  *
59  * If NUMA is not configured or there is only one NUMA node available,
60  * there is no reason to consider NUMA.  This function determines
61  * whether percpu allocation should consider NUMA or not.
62  *
63  * RETURNS:
64  * true if NUMA should be considered; otherwise, false.
65  */
66 static bool __init pcpu_need_numa(void)
67 {
68 #ifdef CONFIG_NEED_MULTIPLE_NODES
69 	pg_data_t *last = NULL;
70 	unsigned int cpu;
71 
72 	for_each_possible_cpu(cpu) {
73 		int node = early_cpu_to_node(cpu);
74 
75 		if (node_online(node) && NODE_DATA(node) &&
76 		    last && last != NODE_DATA(node))
77 			return true;
78 
79 		last = NODE_DATA(node);
80 	}
81 #endif
82 	return false;
83 }
84 #endif
85 
86 /**
87  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
88  * @cpu: cpu to allocate for
89  * @size: size allocation in bytes
90  * @align: alignment
91  *
92  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
93  * does the right thing for NUMA regardless of the current
94  * configuration.
95  *
96  * RETURNS:
97  * Pointer to the allocated area on success, NULL on failure.
98  */
99 static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size,
100 					unsigned long align)
101 {
102 	const unsigned long goal = __pa(MAX_DMA_ADDRESS);
103 #ifdef CONFIG_NEED_MULTIPLE_NODES
104 	int node = early_cpu_to_node(cpu);
105 	void *ptr;
106 
107 	if (!node_online(node) || !NODE_DATA(node)) {
108 		ptr = __alloc_bootmem_nopanic(size, align, goal);
109 		pr_info("cpu %d has no node %d or node-local memory\n",
110 			cpu, node);
111 		pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
112 			 cpu, size, __pa(ptr));
113 	} else {
114 		ptr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
115 						   size, align, goal);
116 		pr_debug("per cpu data for cpu%d %lu bytes on node%d at %016lx\n",
117 			 cpu, size, node, __pa(ptr));
118 	}
119 	return ptr;
120 #else
121 	return __alloc_bootmem_nopanic(size, align, goal);
122 #endif
123 }
124 
125 /*
126  * Helpers for first chunk memory allocation
127  */
128 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
129 {
130 	return pcpu_alloc_bootmem(cpu, size, align);
131 }
132 
133 static void __init pcpu_fc_free(void *ptr, size_t size)
134 {
135 	free_bootmem(__pa(ptr), size);
136 }
137 
138 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
139 {
140 #ifdef CONFIG_NEED_MULTIPLE_NODES
141 	if (early_cpu_to_node(from) == early_cpu_to_node(to))
142 		return LOCAL_DISTANCE;
143 	else
144 		return REMOTE_DISTANCE;
145 #else
146 	return LOCAL_DISTANCE;
147 #endif
148 }
149 
150 static void __init pcpup_populate_pte(unsigned long addr)
151 {
152 	populate_extra_pte(addr);
153 }
154 
155 static inline void setup_percpu_segment(int cpu)
156 {
157 #ifdef CONFIG_X86_32
158 	struct desc_struct gdt;
159 
160 	pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
161 			0x2 | DESCTYPE_S, 0x8);
162 	gdt.s = 1;
163 	write_gdt_entry(get_cpu_gdt_rw(cpu),
164 			GDT_ENTRY_PERCPU, &gdt, 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:%d nr_node_ids:%d\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_warning("%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 		setup_stack_canary_segment(cpu);
228 		/*
229 		 * Copy data used in early init routines from the
230 		 * initial arrays to the per cpu data areas.  These
231 		 * arrays then become expendable and the *_early_ptr's
232 		 * are zeroed indicating that the static arrays are
233 		 * gone.
234 		 */
235 #ifdef CONFIG_X86_LOCAL_APIC
236 		per_cpu(x86_cpu_to_apicid, cpu) =
237 			early_per_cpu_map(x86_cpu_to_apicid, cpu);
238 		per_cpu(x86_bios_cpu_apicid, cpu) =
239 			early_per_cpu_map(x86_bios_cpu_apicid, cpu);
240 		per_cpu(x86_cpu_to_acpiid, cpu) =
241 			early_per_cpu_map(x86_cpu_to_acpiid, cpu);
242 #endif
243 #ifdef CONFIG_X86_32
244 		per_cpu(x86_cpu_to_logical_apicid, cpu) =
245 			early_per_cpu_map(x86_cpu_to_logical_apicid, cpu);
246 #endif
247 #ifdef CONFIG_X86_64
248 		per_cpu(irq_stack_ptr, cpu) =
249 			per_cpu(irq_stack_union.irq_stack, cpu) +
250 			IRQ_STACK_SIZE;
251 #endif
252 #ifdef CONFIG_NUMA
253 		per_cpu(x86_cpu_to_node_map, cpu) =
254 			early_per_cpu_map(x86_cpu_to_node_map, cpu);
255 		/*
256 		 * Ensure that the boot cpu numa_node is correct when the boot
257 		 * cpu is on a node that doesn't have memory installed.
258 		 * Also cpu_up() will call cpu_to_node() for APs when
259 		 * MEMORY_HOTPLUG is defined, before per_cpu(numa_node) is set
260 		 * up later with c_init aka intel_init/amd_init.
261 		 * So set them all (boot cpu and all APs).
262 		 */
263 		set_cpu_numa_node(cpu, early_cpu_to_node(cpu));
264 #endif
265 		/*
266 		 * Up to this point, the boot CPU has been using .init.data
267 		 * area.  Reload any changed state for the boot CPU.
268 		 */
269 		if (!cpu)
270 			switch_to_new_gdt(cpu);
271 	}
272 
273 	/* indicate the early static arrays will soon be gone */
274 #ifdef CONFIG_X86_LOCAL_APIC
275 	early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
276 	early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
277 	early_per_cpu_ptr(x86_cpu_to_acpiid) = NULL;
278 #endif
279 #ifdef CONFIG_X86_32
280 	early_per_cpu_ptr(x86_cpu_to_logical_apicid) = NULL;
281 #endif
282 #ifdef CONFIG_NUMA
283 	early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
284 #endif
285 
286 	/* Setup node to cpumask map */
287 	setup_node_to_cpumask_map();
288 
289 	/* Setup cpu initialized, callin, callout masks */
290 	setup_cpu_local_masks();
291 
292 #ifdef CONFIG_X86_32
293 	/*
294 	 * Sync back kernel address range.  We want to make sure that
295 	 * all kernel mappings, including percpu mappings, are available
296 	 * in the smpboot asm.  We can't reliably pick up percpu
297 	 * mappings using vmalloc_fault(), because exception dispatch
298 	 * needs percpu data.
299 	 */
300 	clone_pgd_range(initial_page_table + KERNEL_PGD_BOUNDARY,
301 			swapper_pg_dir     + KERNEL_PGD_BOUNDARY,
302 			KERNEL_PGD_PTRS);
303 
304 	/*
305 	 * sync back low identity map too.  It is used for example
306 	 * in the 32-bit EFI stub.
307 	 */
308 	clone_pgd_range(initial_page_table,
309 			swapper_pg_dir     + KERNEL_PGD_BOUNDARY,
310 			min(KERNEL_PGD_PTRS, KERNEL_PGD_BOUNDARY));
311 #endif
312 }
313