diff options
Diffstat (limited to 'arch/x86/kernel/setup_percpu.c')
-rw-r--r-- | arch/x86/kernel/setup_percpu.c | 364 |
1 files changed, 45 insertions, 319 deletions
diff --git a/arch/x86/kernel/setup_percpu.c b/arch/x86/kernel/setup_percpu.c index 07d81916f212..d559af913e1f 100644 --- a/arch/x86/kernel/setup_percpu.c +++ b/arch/x86/kernel/setup_percpu.c | |||
@@ -55,6 +55,7 @@ EXPORT_SYMBOL(__per_cpu_offset); | |||
55 | #define PERCPU_FIRST_CHUNK_RESERVE 0 | 55 | #define PERCPU_FIRST_CHUNK_RESERVE 0 |
56 | #endif | 56 | #endif |
57 | 57 | ||
58 | #ifdef CONFIG_X86_32 | ||
58 | /** | 59 | /** |
59 | * pcpu_need_numa - determine percpu allocation needs to consider NUMA | 60 | * pcpu_need_numa - determine percpu allocation needs to consider NUMA |
60 | * | 61 | * |
@@ -83,6 +84,7 @@ static bool __init pcpu_need_numa(void) | |||
83 | #endif | 84 | #endif |
84 | return false; | 85 | return false; |
85 | } | 86 | } |
87 | #endif | ||
86 | 88 | ||
87 | /** | 89 | /** |
88 | * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu | 90 | * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu |
@@ -124,308 +126,35 @@ static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size, | |||
124 | } | 126 | } |
125 | 127 | ||
126 | /* | 128 | /* |
127 | * Large page remap allocator | 129 | * Helpers for first chunk memory allocation |
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 | */ | 130 | */ |
139 | #ifdef CONFIG_NEED_MULTIPLE_NODES | 131 | static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align) |
140 | struct pcpul_ent { | ||
141 | unsigned int cpu; | ||
142 | void *ptr; | ||
143 | }; | ||
144 | |||
145 | static size_t pcpul_size; | ||
146 | static struct pcpul_ent *pcpul_map; | ||
147 | static struct vm_struct pcpul_vm; | ||
148 | |||
149 | static struct page * __init pcpul_get_page(unsigned int cpu, int pageno) | ||
150 | { | 132 | { |
151 | size_t off = (size_t)pageno << PAGE_SHIFT; | 133 | return pcpu_alloc_bootmem(cpu, size, align); |
152 | |||
153 | if (off >= pcpul_size) | ||
154 | return NULL; | ||
155 | |||
156 | return virt_to_page(pcpul_map[cpu].ptr + off); | ||
157 | } | 134 | } |
158 | 135 | ||
159 | static ssize_t __init setup_pcpu_lpage(size_t static_size, bool chosen) | 136 | static void __init pcpu_fc_free(void *ptr, size_t size) |
160 | { | 137 | { |
161 | size_t map_size, dyn_size; | 138 | free_bootmem(__pa(ptr), size); |
162 | unsigned int cpu; | ||
163 | int i, j; | ||
164 | ssize_t ret; | ||
165 | |||
166 | if (!chosen) { | ||
167 | size_t vm_size = VMALLOC_END - VMALLOC_START; | ||
168 | size_t tot_size = nr_cpu_ids * PMD_SIZE; | ||
169 | |||
170 | /* on non-NUMA, embedding is better */ | ||
171 | if (!pcpu_need_numa()) | ||
172 | return -EINVAL; | ||
173 | |||
174 | /* don't consume more than 20% of vmalloc area */ | ||
175 | if (tot_size > vm_size / 5) { | ||
176 | pr_info("PERCPU: too large chunk size %zuMB for " | ||
177 | "large page remap\n", tot_size >> 20); | ||
178 | return -EINVAL; | ||
179 | } | ||
180 | } | ||
181 | |||
182 | /* need PSE */ | ||
183 | if (!cpu_has_pse) { | ||
184 | pr_warning("PERCPU: lpage allocator requires PSE\n"); | ||
185 | return -EINVAL; | ||
186 | } | ||
187 | |||
188 | /* | ||
189 | * Currently supports only single page. Supporting multiple | ||
190 | * pages won't be too difficult if it ever becomes necessary. | ||
191 | */ | ||
192 | pcpul_size = PFN_ALIGN(static_size + PERCPU_MODULE_RESERVE + | ||
193 | PERCPU_DYNAMIC_RESERVE); | ||
194 | if (pcpul_size > PMD_SIZE) { | ||
195 | pr_warning("PERCPU: static data is larger than large page, " | ||
196 | "can't use large page\n"); | ||
197 | return -EINVAL; | ||
198 | } | ||
199 | dyn_size = pcpul_size - static_size - PERCPU_FIRST_CHUNK_RESERVE; | ||
200 | |||
201 | /* allocate pointer array and alloc large pages */ | ||
202 | map_size = PFN_ALIGN(nr_cpu_ids * sizeof(pcpul_map[0])); | ||
203 | pcpul_map = alloc_bootmem(map_size); | ||
204 | |||
205 | for_each_possible_cpu(cpu) { | ||
206 | pcpul_map[cpu].cpu = cpu; | ||
207 | pcpul_map[cpu].ptr = pcpu_alloc_bootmem(cpu, PMD_SIZE, | ||
208 | PMD_SIZE); | ||
209 | if (!pcpul_map[cpu].ptr) { | ||
210 | pr_warning("PERCPU: failed to allocate large page " | ||
211 | "for cpu%u\n", cpu); | ||
212 | goto enomem; | ||
213 | } | ||
214 | |||
215 | /* | ||
216 | * Only use pcpul_size bytes and give back the rest. | ||
217 | * | ||
218 | * Ingo: The 2MB up-rounding bootmem is needed to make | ||
219 | * sure the partial 2MB page is still fully RAM - it's | ||
220 | * not well-specified to have a PAT-incompatible area | ||
221 | * (unmapped RAM, device memory, etc.) in that hole. | ||
222 | */ | ||
223 | free_bootmem(__pa(pcpul_map[cpu].ptr + pcpul_size), | ||
224 | PMD_SIZE - pcpul_size); | ||
225 | |||
226 | memcpy(pcpul_map[cpu].ptr, __per_cpu_load, static_size); | ||
227 | } | ||
228 | |||
229 | /* allocate address and map */ | ||
230 | pcpul_vm.flags = VM_ALLOC; | ||
231 | pcpul_vm.size = nr_cpu_ids * PMD_SIZE; | ||
232 | vm_area_register_early(&pcpul_vm, PMD_SIZE); | ||
233 | |||
234 | for_each_possible_cpu(cpu) { | ||
235 | pmd_t *pmd, pmd_v; | ||
236 | |||
237 | pmd = populate_extra_pmd((unsigned long)pcpul_vm.addr + | ||
238 | cpu * PMD_SIZE); | ||
239 | pmd_v = pfn_pmd(page_to_pfn(virt_to_page(pcpul_map[cpu].ptr)), | ||
240 | PAGE_KERNEL_LARGE); | ||
241 | set_pmd(pmd, pmd_v); | ||
242 | } | ||
243 | |||
244 | /* we're ready, commit */ | ||
245 | pr_info("PERCPU: Remapped at %p with large pages, static data " | ||
246 | "%zu bytes\n", pcpul_vm.addr, static_size); | ||
247 | |||
248 | ret = pcpu_setup_first_chunk(pcpul_get_page, static_size, | ||
249 | PERCPU_FIRST_CHUNK_RESERVE, dyn_size, | ||
250 | PMD_SIZE, pcpul_vm.addr, NULL); | ||
251 | |||
252 | /* sort pcpul_map array for pcpu_lpage_remapped() */ | ||
253 | for (i = 0; i < nr_cpu_ids - 1; i++) | ||
254 | for (j = i + 1; j < nr_cpu_ids; j++) | ||
255 | if (pcpul_map[i].ptr > pcpul_map[j].ptr) { | ||
256 | struct pcpul_ent tmp = pcpul_map[i]; | ||
257 | pcpul_map[i] = pcpul_map[j]; | ||
258 | pcpul_map[j] = tmp; | ||
259 | } | ||
260 | |||
261 | return ret; | ||
262 | |||
263 | enomem: | ||
264 | for_each_possible_cpu(cpu) | ||
265 | if (pcpul_map[cpu].ptr) | ||
266 | free_bootmem(__pa(pcpul_map[cpu].ptr), pcpul_size); | ||
267 | free_bootmem(__pa(pcpul_map), map_size); | ||
268 | return -ENOMEM; | ||
269 | } | 139 | } |
270 | 140 | ||
271 | /** | 141 | static int __init pcpu_cpu_distance(unsigned int from, unsigned int to) |
272 | * pcpu_lpage_remapped - determine whether a kaddr is in pcpul recycled area | ||
273 | * @kaddr: the kernel address in question | ||
274 | * | ||
275 | * Determine whether @kaddr falls in the pcpul recycled area. This is | ||
276 | * used by pageattr to detect VM aliases and break up the pcpu PMD | ||
277 | * mapping such that the same physical page is not mapped under | ||
278 | * different attributes. | ||
279 | * | ||
280 | * The recycled area is always at the tail of a partially used PMD | ||
281 | * page. | ||
282 | * | ||
283 | * RETURNS: | ||
284 | * Address of corresponding remapped pcpu address if match is found; | ||
285 | * otherwise, NULL. | ||
286 | */ | ||
287 | void *pcpu_lpage_remapped(void *kaddr) | ||
288 | { | 142 | { |
289 | void *pmd_addr = (void *)((unsigned long)kaddr & PMD_MASK); | 143 | #ifdef CONFIG_NEED_MULTIPLE_NODES |
290 | unsigned long offset = (unsigned long)kaddr & ~PMD_MASK; | 144 | if (early_cpu_to_node(from) == early_cpu_to_node(to)) |
291 | int left = 0, right = nr_cpu_ids - 1; | 145 | return LOCAL_DISTANCE; |
292 | int pos; | 146 | else |
293 | 147 | return REMOTE_DISTANCE; | |
294 | /* pcpul in use at all? */ | ||
295 | if (!pcpul_map) | ||
296 | return NULL; | ||
297 | |||
298 | /* okay, perform binary search */ | ||
299 | while (left <= right) { | ||
300 | pos = (left + right) / 2; | ||
301 | |||
302 | if (pcpul_map[pos].ptr < pmd_addr) | ||
303 | left = pos + 1; | ||
304 | else if (pcpul_map[pos].ptr > pmd_addr) | ||
305 | right = pos - 1; | ||
306 | else { | ||
307 | /* it shouldn't be in the area for the first chunk */ | ||
308 | WARN_ON(offset < pcpul_size); | ||
309 | |||
310 | return pcpul_vm.addr + | ||
311 | pcpul_map[pos].cpu * PMD_SIZE + offset; | ||
312 | } | ||
313 | } | ||
314 | |||
315 | return NULL; | ||
316 | } | ||
317 | #else | 148 | #else |
318 | static ssize_t __init setup_pcpu_lpage(size_t static_size, bool chosen) | 149 | return LOCAL_DISTANCE; |
319 | { | ||
320 | return -EINVAL; | ||
321 | } | ||
322 | #endif | 150 | #endif |
323 | |||
324 | /* | ||
325 | * Embedding allocator | ||
326 | * | ||
327 | * The first chunk is sized to just contain the static area plus | ||
328 | * module and dynamic reserves and embedded into linear physical | ||
329 | * mapping so that it can use PMD mapping without additional TLB | ||
330 | * pressure. | ||
331 | */ | ||
332 | static ssize_t __init setup_pcpu_embed(size_t static_size, bool chosen) | ||
333 | { | ||
334 | size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE; | ||
335 | |||
336 | /* | ||
337 | * If large page isn't supported, there's no benefit in doing | ||
338 | * this. Also, embedding allocation doesn't play well with | ||
339 | * NUMA. | ||
340 | */ | ||
341 | if (!chosen && (!cpu_has_pse || pcpu_need_numa())) | ||
342 | return -EINVAL; | ||
343 | |||
344 | return pcpu_embed_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE, | ||
345 | reserve - PERCPU_FIRST_CHUNK_RESERVE, -1); | ||
346 | } | 151 | } |
347 | 152 | ||
348 | /* | 153 | static void __init pcpup_populate_pte(unsigned long addr) |
349 | * 4k page allocator | ||
350 | * | ||
351 | * This is the basic allocator. Static percpu area is allocated | ||
352 | * page-by-page and most of initialization is done by the generic | ||
353 | * setup function. | ||
354 | */ | ||
355 | static struct page **pcpu4k_pages __initdata; | ||
356 | static int pcpu4k_nr_static_pages __initdata; | ||
357 | |||
358 | static struct page * __init pcpu4k_get_page(unsigned int cpu, int pageno) | ||
359 | { | ||
360 | if (pageno < pcpu4k_nr_static_pages) | ||
361 | return pcpu4k_pages[cpu * pcpu4k_nr_static_pages + pageno]; | ||
362 | return NULL; | ||
363 | } | ||
364 | |||
365 | static void __init pcpu4k_populate_pte(unsigned long addr) | ||
366 | { | 154 | { |
367 | populate_extra_pte(addr); | 155 | populate_extra_pte(addr); |
368 | } | 156 | } |
369 | 157 | ||
370 | static ssize_t __init setup_pcpu_4k(size_t static_size) | ||
371 | { | ||
372 | size_t pages_size; | ||
373 | unsigned int cpu; | ||
374 | int i, j; | ||
375 | ssize_t ret; | ||
376 | |||
377 | pcpu4k_nr_static_pages = PFN_UP(static_size); | ||
378 | |||
379 | /* unaligned allocations can't be freed, round up to page size */ | ||
380 | pages_size = PFN_ALIGN(pcpu4k_nr_static_pages * nr_cpu_ids | ||
381 | * sizeof(pcpu4k_pages[0])); | ||
382 | pcpu4k_pages = alloc_bootmem(pages_size); | ||
383 | |||
384 | /* allocate and copy */ | ||
385 | j = 0; | ||
386 | for_each_possible_cpu(cpu) | ||
387 | for (i = 0; i < pcpu4k_nr_static_pages; i++) { | ||
388 | void *ptr; | ||
389 | |||
390 | ptr = pcpu_alloc_bootmem(cpu, PAGE_SIZE, PAGE_SIZE); | ||
391 | if (!ptr) { | ||
392 | pr_warning("PERCPU: failed to allocate " | ||
393 | "4k page for cpu%u\n", cpu); | ||
394 | goto enomem; | ||
395 | } | ||
396 | |||
397 | memcpy(ptr, __per_cpu_load + i * PAGE_SIZE, PAGE_SIZE); | ||
398 | pcpu4k_pages[j++] = virt_to_page(ptr); | ||
399 | } | ||
400 | |||
401 | /* we're ready, commit */ | ||
402 | pr_info("PERCPU: Allocated %d 4k pages, static data %zu bytes\n", | ||
403 | pcpu4k_nr_static_pages, static_size); | ||
404 | |||
405 | ret = pcpu_setup_first_chunk(pcpu4k_get_page, static_size, | ||
406 | PERCPU_FIRST_CHUNK_RESERVE, -1, | ||
407 | -1, NULL, pcpu4k_populate_pte); | ||
408 | goto out_free_ar; | ||
409 | |||
410 | enomem: | ||
411 | while (--j >= 0) | ||
412 | free_bootmem(__pa(page_address(pcpu4k_pages[j])), PAGE_SIZE); | ||
413 | ret = -ENOMEM; | ||
414 | out_free_ar: | ||
415 | free_bootmem(__pa(pcpu4k_pages), pages_size); | ||
416 | return ret; | ||
417 | } | ||
418 | |||
419 | /* for explicit first chunk allocator selection */ | ||
420 | static char pcpu_chosen_alloc[16] __initdata; | ||
421 | |||
422 | static int __init percpu_alloc_setup(char *str) | ||
423 | { | ||
424 | strncpy(pcpu_chosen_alloc, str, sizeof(pcpu_chosen_alloc) - 1); | ||
425 | return 0; | ||
426 | } | ||
427 | early_param("percpu_alloc", percpu_alloc_setup); | ||
428 | |||
429 | static inline void setup_percpu_segment(int cpu) | 158 | static inline void setup_percpu_segment(int cpu) |
430 | { | 159 | { |
431 | #ifdef CONFIG_X86_32 | 160 | #ifdef CONFIG_X86_32 |
@@ -441,52 +170,49 @@ static inline void setup_percpu_segment(int cpu) | |||
441 | 170 | ||
442 | void __init setup_per_cpu_areas(void) | 171 | void __init setup_per_cpu_areas(void) |
443 | { | 172 | { |
444 | size_t static_size = __per_cpu_end - __per_cpu_start; | ||
445 | unsigned int cpu; | 173 | unsigned int cpu; |
446 | unsigned long delta; | 174 | unsigned long delta; |
447 | size_t pcpu_unit_size; | 175 | int rc; |
448 | ssize_t ret; | ||
449 | 176 | ||
450 | pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n", | 177 | pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n", |
451 | NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids); | 178 | NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids); |
452 | 179 | ||
453 | /* | 180 | /* |
454 | * Allocate percpu area. If PSE is supported, try to make use | 181 | * Allocate percpu area. Embedding allocator is our favorite; |
455 | * of large page mappings. Please read comments on top of | 182 | * however, on NUMA configurations, it can result in very |
456 | * each allocator for details. | 183 | * sparse unit mapping and vmalloc area isn't spacious enough |
184 | * on 32bit. Use page in that case. | ||
457 | */ | 185 | */ |
458 | ret = -EINVAL; | 186 | #ifdef CONFIG_X86_32 |
459 | if (strlen(pcpu_chosen_alloc)) { | 187 | if (pcpu_chosen_fc == PCPU_FC_AUTO && pcpu_need_numa()) |
460 | if (strcmp(pcpu_chosen_alloc, "4k")) { | 188 | pcpu_chosen_fc = PCPU_FC_PAGE; |
461 | if (!strcmp(pcpu_chosen_alloc, "lpage")) | 189 | #endif |
462 | ret = setup_pcpu_lpage(static_size, true); | 190 | rc = -EINVAL; |
463 | else if (!strcmp(pcpu_chosen_alloc, "embed")) | 191 | if (pcpu_chosen_fc != PCPU_FC_PAGE) { |
464 | ret = setup_pcpu_embed(static_size, true); | 192 | const size_t atom_size = cpu_has_pse ? PMD_SIZE : PAGE_SIZE; |
465 | else | 193 | const size_t dyn_size = PERCPU_MODULE_RESERVE + |
466 | pr_warning("PERCPU: unknown allocator %s " | 194 | PERCPU_DYNAMIC_RESERVE - PERCPU_FIRST_CHUNK_RESERVE; |
467 | "specified\n", pcpu_chosen_alloc); | 195 | |
468 | if (ret < 0) | 196 | rc = pcpu_embed_first_chunk(PERCPU_FIRST_CHUNK_RESERVE, |
469 | pr_warning("PERCPU: %s allocator failed (%zd), " | 197 | dyn_size, atom_size, |
470 | "falling back to 4k\n", | 198 | pcpu_cpu_distance, |
471 | pcpu_chosen_alloc, ret); | 199 | pcpu_fc_alloc, pcpu_fc_free); |
472 | } | 200 | if (rc < 0) |
473 | } else { | 201 | pr_warning("PERCPU: %s allocator failed (%d), " |
474 | ret = setup_pcpu_lpage(static_size, false); | 202 | "falling back to page size\n", |
475 | if (ret < 0) | 203 | pcpu_fc_names[pcpu_chosen_fc], rc); |
476 | ret = setup_pcpu_embed(static_size, false); | ||
477 | } | 204 | } |
478 | if (ret < 0) | 205 | if (rc < 0) |
479 | ret = setup_pcpu_4k(static_size); | 206 | rc = pcpu_page_first_chunk(PERCPU_FIRST_CHUNK_RESERVE, |
480 | if (ret < 0) | 207 | pcpu_fc_alloc, pcpu_fc_free, |
481 | panic("cannot allocate static percpu area (%zu bytes, err=%zd)", | 208 | pcpup_populate_pte); |
482 | static_size, ret); | 209 | if (rc < 0) |
483 | 210 | panic("cannot initialize percpu area (err=%d)", rc); | |
484 | pcpu_unit_size = ret; | ||
485 | 211 | ||
486 | /* alrighty, percpu areas up and running */ | 212 | /* alrighty, percpu areas up and running */ |
487 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | 213 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; |
488 | for_each_possible_cpu(cpu) { | 214 | for_each_possible_cpu(cpu) { |
489 | per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size; | 215 | per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu]; |
490 | per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu); | 216 | per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu); |
491 | per_cpu(cpu_number, cpu) = cpu; | 217 | per_cpu(cpu_number, cpu) = cpu; |
492 | setup_percpu_segment(cpu); | 218 | setup_percpu_segment(cpu); |