Merge git://git.kernel.org/pub/scm/linux/kernel/git/arjan/linux-2.6-async-for-30
[linux-2.6] / arch / x86 / kernel / setup_percpu.c
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>
8 #include <linux/smp.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>
19 #include <asm/cpu.h>
20 #include <asm/stackprotector.h>
21
22 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
23 # define DBG(x...) printk(KERN_DEBUG x)
24 #else
25 # define DBG(x...)
26 #endif
27
28 DEFINE_PER_CPU(int, cpu_number);
29 EXPORT_PER_CPU_SYMBOL(cpu_number);
30
31 #ifdef CONFIG_X86_64
32 #define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
33 #else
34 #define BOOT_PERCPU_OFFSET 0
35 #endif
36
37 DEFINE_PER_CPU(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
38 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
39
40 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly = {
41         [0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
42 };
43 EXPORT_SYMBOL(__per_cpu_offset);
44
45 /*
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.
51  */
52 #ifdef CONFIG_X86_64
53 #define PERCPU_FIRST_CHUNK_RESERVE      PERCPU_MODULE_RESERVE
54 #else
55 #define PERCPU_FIRST_CHUNK_RESERVE      0
56 #endif
57
58 /**
59  * pcpu_need_numa - determine percpu allocation needs to consider NUMA
60  *
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.
64  *
65  * RETURNS:
66  * true if NUMA should be considered; otherwise, false.
67  */
68 static bool __init pcpu_need_numa(void)
69 {
70 #ifdef CONFIG_NEED_MULTIPLE_NODES
71         pg_data_t *last = NULL;
72         unsigned int cpu;
73
74         for_each_possible_cpu(cpu) {
75                 int node = early_cpu_to_node(cpu);
76
77                 if (node_online(node) && NODE_DATA(node) &&
78                     last && last != NODE_DATA(node))
79                         return true;
80
81                 last = NODE_DATA(node);
82         }
83 #endif
84         return false;
85 }
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_NEED_MULTIPLE_NODES
105         int node = early_cpu_to_node(cpu);
106         void *ptr;
107
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",
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 = __alloc_bootmem_node_nopanic(NODE_DATA(node),
116                                                    size, align, goal);
117                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
118                          "%016lx\n", cpu, size, node, __pa(ptr));
119         }
120         return ptr;
121 #else
122         return __alloc_bootmem_nopanic(size, align, goal);
123 #endif
124 }
125
126 /*
127  * Remap allocator
128  *
129  * This allocator uses PMD page as unit.  A PMD page is allocated for
130  * each cpu and each is remapped into vmalloc area using PMD mapping.
131  * As PMD page is quite large, only part of it is used for the first
132  * chunk.  Unused part is returned to the bootmem allocator.
133  *
134  * So, the PMD pages are mapped twice - once to the physical mapping
135  * and to the vmalloc area for the first percpu chunk.  The double
136  * mapping does add one more PMD TLB entry pressure but still is much
137  * better than only using 4k mappings while still being NUMA friendly.
138  */
139 #ifdef CONFIG_NEED_MULTIPLE_NODES
140 static size_t pcpur_size __initdata;
141 static void **pcpur_ptrs __initdata;
142
143 static struct page * __init pcpur_get_page(unsigned int cpu, int pageno)
144 {
145         size_t off = (size_t)pageno << PAGE_SHIFT;
146
147         if (off >= pcpur_size)
148                 return NULL;
149
150         return virt_to_page(pcpur_ptrs[cpu] + off);
151 }
152
153 static ssize_t __init setup_pcpu_remap(size_t static_size)
154 {
155         static struct vm_struct vm;
156         pg_data_t *last;
157         size_t ptrs_size, dyn_size;
158         unsigned int cpu;
159         ssize_t ret;
160
161         /*
162          * If large page isn't supported, there's no benefit in doing
163          * this.  Also, on non-NUMA, embedding is better.
164          */
165         if (!cpu_has_pse || pcpu_need_numa())
166                 return -EINVAL;
167
168         last = NULL;
169         for_each_possible_cpu(cpu) {
170                 int node = early_cpu_to_node(cpu);
171
172                 if (node_online(node) && NODE_DATA(node) &&
173                     last && last != NODE_DATA(node))
174                         goto proceed;
175
176                 last = NODE_DATA(node);
177         }
178         return -EINVAL;
179
180 proceed:
181         /*
182          * Currently supports only single page.  Supporting multiple
183          * pages won't be too difficult if it ever becomes necessary.
184          */
185         pcpur_size = PFN_ALIGN(static_size + PERCPU_MODULE_RESERVE +
186                                PERCPU_DYNAMIC_RESERVE);
187         if (pcpur_size > PMD_SIZE) {
188                 pr_warning("PERCPU: static data is larger than large page, "
189                            "can't use large page\n");
190                 return -EINVAL;
191         }
192         dyn_size = pcpur_size - static_size - PERCPU_FIRST_CHUNK_RESERVE;
193
194         /* allocate pointer array and alloc large pages */
195         ptrs_size = PFN_ALIGN(num_possible_cpus() * sizeof(pcpur_ptrs[0]));
196         pcpur_ptrs = alloc_bootmem(ptrs_size);
197
198         for_each_possible_cpu(cpu) {
199                 pcpur_ptrs[cpu] = pcpu_alloc_bootmem(cpu, PMD_SIZE, PMD_SIZE);
200                 if (!pcpur_ptrs[cpu])
201                         goto enomem;
202
203                 /*
204                  * Only use pcpur_size bytes and give back the rest.
205                  *
206                  * Ingo: The 2MB up-rounding bootmem is needed to make
207                  * sure the partial 2MB page is still fully RAM - it's
208                  * not well-specified to have a PAT-incompatible area
209                  * (unmapped RAM, device memory, etc.) in that hole.
210                  */
211                 free_bootmem(__pa(pcpur_ptrs[cpu] + pcpur_size),
212                              PMD_SIZE - pcpur_size);
213
214                 memcpy(pcpur_ptrs[cpu], __per_cpu_load, static_size);
215         }
216
217         /* allocate address and map */
218         vm.flags = VM_ALLOC;
219         vm.size = num_possible_cpus() * PMD_SIZE;
220         vm_area_register_early(&vm, PMD_SIZE);
221
222         for_each_possible_cpu(cpu) {
223                 pmd_t *pmd;
224
225                 pmd = populate_extra_pmd((unsigned long)vm.addr
226                                          + cpu * PMD_SIZE);
227                 set_pmd(pmd, pfn_pmd(page_to_pfn(virt_to_page(pcpur_ptrs[cpu])),
228                                      PAGE_KERNEL_LARGE));
229         }
230
231         /* we're ready, commit */
232         pr_info("PERCPU: Remapped at %p with large pages, static data "
233                 "%zu bytes\n", vm.addr, static_size);
234
235         ret = pcpu_setup_first_chunk(pcpur_get_page, static_size,
236                                      PERCPU_FIRST_CHUNK_RESERVE, dyn_size,
237                                      PMD_SIZE, vm.addr, NULL);
238         goto out_free_ar;
239
240 enomem:
241         for_each_possible_cpu(cpu)
242                 if (pcpur_ptrs[cpu])
243                         free_bootmem(__pa(pcpur_ptrs[cpu]), PMD_SIZE);
244         ret = -ENOMEM;
245 out_free_ar:
246         free_bootmem(__pa(pcpur_ptrs), ptrs_size);
247         return ret;
248 }
249 #else
250 static ssize_t __init setup_pcpu_remap(size_t static_size)
251 {
252         return -EINVAL;
253 }
254 #endif
255
256 /*
257  * Embedding allocator
258  *
259  * The first chunk is sized to just contain the static area plus
260  * module and dynamic reserves and embedded into linear physical
261  * mapping so that it can use PMD mapping without additional TLB
262  * pressure.
263  */
264 static ssize_t __init setup_pcpu_embed(size_t static_size)
265 {
266         size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
267
268         /*
269          * If large page isn't supported, there's no benefit in doing
270          * this.  Also, embedding allocation doesn't play well with
271          * NUMA.
272          */
273         if (!cpu_has_pse || pcpu_need_numa())
274                 return -EINVAL;
275
276         return pcpu_embed_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE,
277                                       reserve - PERCPU_FIRST_CHUNK_RESERVE, -1);
278 }
279
280 /*
281  * 4k page allocator
282  *
283  * This is the basic allocator.  Static percpu area is allocated
284  * page-by-page and most of initialization is done by the generic
285  * setup function.
286  */
287 static struct page **pcpu4k_pages __initdata;
288 static int pcpu4k_nr_static_pages __initdata;
289
290 static struct page * __init pcpu4k_get_page(unsigned int cpu, int pageno)
291 {
292         if (pageno < pcpu4k_nr_static_pages)
293                 return pcpu4k_pages[cpu * pcpu4k_nr_static_pages + pageno];
294         return NULL;
295 }
296
297 static void __init pcpu4k_populate_pte(unsigned long addr)
298 {
299         populate_extra_pte(addr);
300 }
301
302 static ssize_t __init setup_pcpu_4k(size_t static_size)
303 {
304         size_t pages_size;
305         unsigned int cpu;
306         int i, j;
307         ssize_t ret;
308
309         pcpu4k_nr_static_pages = PFN_UP(static_size);
310
311         /* unaligned allocations can't be freed, round up to page size */
312         pages_size = PFN_ALIGN(pcpu4k_nr_static_pages * num_possible_cpus()
313                                * sizeof(pcpu4k_pages[0]));
314         pcpu4k_pages = alloc_bootmem(pages_size);
315
316         /* allocate and copy */
317         j = 0;
318         for_each_possible_cpu(cpu)
319                 for (i = 0; i < pcpu4k_nr_static_pages; i++) {
320                         void *ptr;
321
322                         ptr = pcpu_alloc_bootmem(cpu, PAGE_SIZE, PAGE_SIZE);
323                         if (!ptr)
324                                 goto enomem;
325
326                         memcpy(ptr, __per_cpu_load + i * PAGE_SIZE, PAGE_SIZE);
327                         pcpu4k_pages[j++] = virt_to_page(ptr);
328                 }
329
330         /* we're ready, commit */
331         pr_info("PERCPU: Allocated %d 4k pages, static data %zu bytes\n",
332                 pcpu4k_nr_static_pages, static_size);
333
334         ret = pcpu_setup_first_chunk(pcpu4k_get_page, static_size,
335                                      PERCPU_FIRST_CHUNK_RESERVE, -1,
336                                      -1, NULL, pcpu4k_populate_pte);
337         goto out_free_ar;
338
339 enomem:
340         while (--j >= 0)
341                 free_bootmem(__pa(page_address(pcpu4k_pages[j])), PAGE_SIZE);
342         ret = -ENOMEM;
343 out_free_ar:
344         free_bootmem(__pa(pcpu4k_pages), pages_size);
345         return ret;
346 }
347
348 static inline void setup_percpu_segment(int cpu)
349 {
350 #ifdef CONFIG_X86_32
351         struct desc_struct gdt;
352
353         pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
354                         0x2 | DESCTYPE_S, 0x8);
355         gdt.s = 1;
356         write_gdt_entry(get_cpu_gdt_table(cpu),
357                         GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S);
358 #endif
359 }
360
361 /*
362  * Great future plan:
363  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
364  * Always point %gs to its beginning
365  */
366 void __init setup_per_cpu_areas(void)
367 {
368         size_t static_size = __per_cpu_end - __per_cpu_start;
369         unsigned int cpu;
370         unsigned long delta;
371         size_t pcpu_unit_size;
372         ssize_t ret;
373
374         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
375                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
376
377         /*
378          * Allocate percpu area.  If PSE is supported, try to make use
379          * of large page mappings.  Please read comments on top of
380          * each allocator for details.
381          */
382         ret = setup_pcpu_remap(static_size);
383         if (ret < 0)
384                 ret = setup_pcpu_embed(static_size);
385         if (ret < 0)
386                 ret = setup_pcpu_4k(static_size);
387         if (ret < 0)
388                 panic("cannot allocate static percpu area (%zu bytes, err=%zd)",
389                       static_size, ret);
390
391         pcpu_unit_size = ret;
392
393         /* alrighty, percpu areas up and running */
394         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
395         for_each_possible_cpu(cpu) {
396                 per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size;
397                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
398                 per_cpu(cpu_number, cpu) = cpu;
399                 setup_percpu_segment(cpu);
400                 setup_stack_canary_segment(cpu);
401                 /*
402                  * Copy data used in early init routines from the
403                  * initial arrays to the per cpu data areas.  These
404                  * arrays then become expendable and the *_early_ptr's
405                  * are zeroed indicating that the static arrays are
406                  * gone.
407                  */
408 #ifdef CONFIG_X86_LOCAL_APIC
409                 per_cpu(x86_cpu_to_apicid, cpu) =
410                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
411                 per_cpu(x86_bios_cpu_apicid, cpu) =
412                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
413 #endif
414 #ifdef CONFIG_X86_64
415                 per_cpu(irq_stack_ptr, cpu) =
416                         per_cpu(irq_stack_union.irq_stack, cpu) +
417                         IRQ_STACK_SIZE - 64;
418 #ifdef CONFIG_NUMA
419                 per_cpu(x86_cpu_to_node_map, cpu) =
420                         early_per_cpu_map(x86_cpu_to_node_map, cpu);
421 #endif
422 #endif
423                 /*
424                  * Up to this point, the boot CPU has been using .data.init
425                  * area.  Reload any changed state for the boot CPU.
426                  */
427                 if (cpu == boot_cpu_id)
428                         switch_to_new_gdt(cpu);
429         }
430
431         /* indicate the early static arrays will soon be gone */
432 #ifdef CONFIG_X86_LOCAL_APIC
433         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
434         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
435 #endif
436 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
437         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
438 #endif
439
440         /* Setup node to cpumask map */
441         setup_node_to_cpumask_map();
442
443         /* Setup cpu initialized, callin, callout masks */
444         setup_cpu_local_masks();
445 }