percpu.c revision e933a73f48e3b2d40cfa56d81e2646f194b5a66a
1/* 2 * linux/mm/percpu.c - percpu memory allocator 3 * 4 * Copyright (C) 2009 SUSE Linux Products GmbH 5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org> 6 * 7 * This file is released under the GPLv2. 8 * 9 * This is percpu allocator which can handle both static and dynamic 10 * areas. Percpu areas are allocated in chunks in vmalloc area. Each 11 * chunk is consisted of boot-time determined number of units and the 12 * first chunk is used for static percpu variables in the kernel image 13 * (special boot time alloc/init handling necessary as these areas 14 * need to be brought up before allocation services are running). 15 * Unit grows as necessary and all units grow or shrink in unison. 16 * When a chunk is filled up, another chunk is allocated. ie. in 17 * vmalloc area 18 * 19 * c0 c1 c2 20 * ------------------- ------------------- ------------ 21 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u 22 * ------------------- ...... ------------------- .... ------------ 23 * 24 * Allocation is done in offset-size areas of single unit space. Ie, 25 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0, 26 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to 27 * cpus. On NUMA, the mapping can be non-linear and even sparse. 28 * Percpu access can be done by configuring percpu base registers 29 * according to cpu to unit mapping and pcpu_unit_size. 30 * 31 * There are usually many small percpu allocations many of them being 32 * as small as 4 bytes. The allocator organizes chunks into lists 33 * according to free size and tries to allocate from the fullest one. 34 * Each chunk keeps the maximum contiguous area size hint which is 35 * guaranteed to be eqaul to or larger than the maximum contiguous 36 * area in the chunk. This helps the allocator not to iterate the 37 * chunk maps unnecessarily. 38 * 39 * Allocation state in each chunk is kept using an array of integers 40 * on chunk->map. A positive value in the map represents a free 41 * region and negative allocated. Allocation inside a chunk is done 42 * by scanning this map sequentially and serving the first matching 43 * entry. This is mostly copied from the percpu_modalloc() allocator. 44 * Chunks can be determined from the address using the index field 45 * in the page struct. The index field contains a pointer to the chunk. 46 * 47 * To use this allocator, arch code should do the followings. 48 * 49 * - drop CONFIG_HAVE_LEGACY_PER_CPU_AREA 50 * 51 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate 52 * regular address to percpu pointer and back if they need to be 53 * different from the default 54 * 55 * - use pcpu_setup_first_chunk() during percpu area initialization to 56 * setup the first chunk containing the kernel static percpu area 57 */ 58 59#include <linux/bitmap.h> 60#include <linux/bootmem.h> 61#include <linux/err.h> 62#include <linux/list.h> 63#include <linux/log2.h> 64#include <linux/mm.h> 65#include <linux/module.h> 66#include <linux/mutex.h> 67#include <linux/percpu.h> 68#include <linux/pfn.h> 69#include <linux/slab.h> 70#include <linux/spinlock.h> 71#include <linux/vmalloc.h> 72#include <linux/workqueue.h> 73 74#include <asm/cacheflush.h> 75#include <asm/sections.h> 76#include <asm/tlbflush.h> 77 78#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ 79#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ 80 81/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ 82#ifndef __addr_to_pcpu_ptr 83#define __addr_to_pcpu_ptr(addr) \ 84 (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \ 85 + (unsigned long)__per_cpu_start) 86#endif 87#ifndef __pcpu_ptr_to_addr 88#define __pcpu_ptr_to_addr(ptr) \ 89 (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \ 90 - (unsigned long)__per_cpu_start) 91#endif 92 93struct pcpu_chunk { 94 struct list_head list; /* linked to pcpu_slot lists */ 95 int free_size; /* free bytes in the chunk */ 96 int contig_hint; /* max contiguous size hint */ 97 void *base_addr; /* base address of this chunk */ 98 int map_used; /* # of map entries used */ 99 int map_alloc; /* # of map entries allocated */ 100 int *map; /* allocation map */ 101 struct vm_struct **vms; /* mapped vmalloc regions */ 102 bool immutable; /* no [de]population allowed */ 103 unsigned long populated[]; /* populated bitmap */ 104}; 105 106static int pcpu_unit_pages __read_mostly; 107static int pcpu_unit_size __read_mostly; 108static int pcpu_nr_units __read_mostly; 109static int pcpu_atom_size __read_mostly; 110static int pcpu_nr_slots __read_mostly; 111static size_t pcpu_chunk_struct_size __read_mostly; 112 113/* cpus with the lowest and highest unit numbers */ 114static unsigned int pcpu_first_unit_cpu __read_mostly; 115static unsigned int pcpu_last_unit_cpu __read_mostly; 116 117/* the address of the first chunk which starts with the kernel static area */ 118void *pcpu_base_addr __read_mostly; 119EXPORT_SYMBOL_GPL(pcpu_base_addr); 120 121static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */ 122const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */ 123 124/* group information, used for vm allocation */ 125static int pcpu_nr_groups __read_mostly; 126static const unsigned long *pcpu_group_offsets __read_mostly; 127static const size_t *pcpu_group_sizes __read_mostly; 128 129/* 130 * The first chunk which always exists. Note that unlike other 131 * chunks, this one can be allocated and mapped in several different 132 * ways and thus often doesn't live in the vmalloc area. 133 */ 134static struct pcpu_chunk *pcpu_first_chunk; 135 136/* 137 * Optional reserved chunk. This chunk reserves part of the first 138 * chunk and serves it for reserved allocations. The amount of 139 * reserved offset is in pcpu_reserved_chunk_limit. When reserved 140 * area doesn't exist, the following variables contain NULL and 0 141 * respectively. 142 */ 143static struct pcpu_chunk *pcpu_reserved_chunk; 144static int pcpu_reserved_chunk_limit; 145 146/* 147 * Synchronization rules. 148 * 149 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former 150 * protects allocation/reclaim paths, chunks, populated bitmap and 151 * vmalloc mapping. The latter is a spinlock and protects the index 152 * data structures - chunk slots, chunks and area maps in chunks. 153 * 154 * During allocation, pcpu_alloc_mutex is kept locked all the time and 155 * pcpu_lock is grabbed and released as necessary. All actual memory 156 * allocations are done using GFP_KERNEL with pcpu_lock released. 157 * 158 * Free path accesses and alters only the index data structures, so it 159 * can be safely called from atomic context. When memory needs to be 160 * returned to the system, free path schedules reclaim_work which 161 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be 162 * reclaimed, release both locks and frees the chunks. Note that it's 163 * necessary to grab both locks to remove a chunk from circulation as 164 * allocation path might be referencing the chunk with only 165 * pcpu_alloc_mutex locked. 166 */ 167static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ 168static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ 169 170static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ 171 172/* reclaim work to release fully free chunks, scheduled from free path */ 173static void pcpu_reclaim(struct work_struct *work); 174static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); 175 176static int __pcpu_size_to_slot(int size) 177{ 178 int highbit = fls(size); /* size is in bytes */ 179 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 180} 181 182static int pcpu_size_to_slot(int size) 183{ 184 if (size == pcpu_unit_size) 185 return pcpu_nr_slots - 1; 186 return __pcpu_size_to_slot(size); 187} 188 189static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 190{ 191 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) 192 return 0; 193 194 return pcpu_size_to_slot(chunk->free_size); 195} 196 197static int pcpu_page_idx(unsigned int cpu, int page_idx) 198{ 199 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 200} 201 202static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 203 unsigned int cpu, int page_idx) 204{ 205 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] + 206 (page_idx << PAGE_SHIFT); 207} 208 209static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk, 210 unsigned int cpu, int page_idx) 211{ 212 /* must not be used on pre-mapped chunk */ 213 WARN_ON(chunk->immutable); 214 215 return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx)); 216} 217 218/* set the pointer to a chunk in a page struct */ 219static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 220{ 221 page->index = (unsigned long)pcpu; 222} 223 224/* obtain pointer to a chunk from a page struct */ 225static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 226{ 227 return (struct pcpu_chunk *)page->index; 228} 229 230static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end) 231{ 232 *rs = find_next_zero_bit(chunk->populated, end, *rs); 233 *re = find_next_bit(chunk->populated, end, *rs + 1); 234} 235 236static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end) 237{ 238 *rs = find_next_bit(chunk->populated, end, *rs); 239 *re = find_next_zero_bit(chunk->populated, end, *rs + 1); 240} 241 242/* 243 * (Un)populated page region iterators. Iterate over (un)populated 244 * page regions betwen @start and @end in @chunk. @rs and @re should 245 * be integer variables and will be set to start and end page index of 246 * the current region. 247 */ 248#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ 249 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ 250 (rs) < (re); \ 251 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) 252 253#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ 254 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ 255 (rs) < (re); \ 256 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) 257 258/** 259 * pcpu_mem_alloc - allocate memory 260 * @size: bytes to allocate 261 * 262 * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 263 * kzalloc() is used; otherwise, vmalloc() is used. The returned 264 * memory is always zeroed. 265 * 266 * CONTEXT: 267 * Does GFP_KERNEL allocation. 268 * 269 * RETURNS: 270 * Pointer to the allocated area on success, NULL on failure. 271 */ 272static void *pcpu_mem_alloc(size_t size) 273{ 274 if (size <= PAGE_SIZE) 275 return kzalloc(size, GFP_KERNEL); 276 else { 277 void *ptr = vmalloc(size); 278 if (ptr) 279 memset(ptr, 0, size); 280 return ptr; 281 } 282} 283 284/** 285 * pcpu_mem_free - free memory 286 * @ptr: memory to free 287 * @size: size of the area 288 * 289 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). 290 */ 291static void pcpu_mem_free(void *ptr, size_t size) 292{ 293 if (size <= PAGE_SIZE) 294 kfree(ptr); 295 else 296 vfree(ptr); 297} 298 299/** 300 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 301 * @chunk: chunk of interest 302 * @oslot: the previous slot it was on 303 * 304 * This function is called after an allocation or free changed @chunk. 305 * New slot according to the changed state is determined and @chunk is 306 * moved to the slot. Note that the reserved chunk is never put on 307 * chunk slots. 308 * 309 * CONTEXT: 310 * pcpu_lock. 311 */ 312static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 313{ 314 int nslot = pcpu_chunk_slot(chunk); 315 316 if (chunk != pcpu_reserved_chunk && oslot != nslot) { 317 if (oslot < nslot) 318 list_move(&chunk->list, &pcpu_slot[nslot]); 319 else 320 list_move_tail(&chunk->list, &pcpu_slot[nslot]); 321 } 322} 323 324/** 325 * pcpu_chunk_addr_search - determine chunk containing specified address 326 * @addr: address for which the chunk needs to be determined. 327 * 328 * RETURNS: 329 * The address of the found chunk. 330 */ 331static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 332{ 333 void *first_start = pcpu_first_chunk->base_addr; 334 335 /* is it in the first chunk? */ 336 if (addr >= first_start && addr < first_start + pcpu_unit_size) { 337 /* is it in the reserved area? */ 338 if (addr < first_start + pcpu_reserved_chunk_limit) 339 return pcpu_reserved_chunk; 340 return pcpu_first_chunk; 341 } 342 343 /* 344 * The address is relative to unit0 which might be unused and 345 * thus unmapped. Offset the address to the unit space of the 346 * current processor before looking it up in the vmalloc 347 * space. Note that any possible cpu id can be used here, so 348 * there's no need to worry about preemption or cpu hotplug. 349 */ 350 addr += pcpu_unit_offsets[smp_processor_id()]; 351 return pcpu_get_page_chunk(vmalloc_to_page(addr)); 352} 353 354/** 355 * pcpu_extend_area_map - extend area map for allocation 356 * @chunk: target chunk 357 * 358 * Extend area map of @chunk so that it can accomodate an allocation. 359 * A single allocation can split an area into three areas, so this 360 * function makes sure that @chunk->map has at least two extra slots. 361 * 362 * CONTEXT: 363 * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired 364 * if area map is extended. 365 * 366 * RETURNS: 367 * 0 if noop, 1 if successfully extended, -errno on failure. 368 */ 369static int pcpu_extend_area_map(struct pcpu_chunk *chunk) 370{ 371 int new_alloc; 372 int *new; 373 size_t size; 374 375 /* has enough? */ 376 if (chunk->map_alloc >= chunk->map_used + 2) 377 return 0; 378 379 spin_unlock_irq(&pcpu_lock); 380 381 new_alloc = PCPU_DFL_MAP_ALLOC; 382 while (new_alloc < chunk->map_used + 2) 383 new_alloc *= 2; 384 385 new = pcpu_mem_alloc(new_alloc * sizeof(new[0])); 386 if (!new) { 387 spin_lock_irq(&pcpu_lock); 388 return -ENOMEM; 389 } 390 391 /* 392 * Acquire pcpu_lock and switch to new area map. Only free 393 * could have happened inbetween, so map_used couldn't have 394 * grown. 395 */ 396 spin_lock_irq(&pcpu_lock); 397 BUG_ON(new_alloc < chunk->map_used + 2); 398 399 size = chunk->map_alloc * sizeof(chunk->map[0]); 400 memcpy(new, chunk->map, size); 401 402 /* 403 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is 404 * one of the first chunks and still using static map. 405 */ 406 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC) 407 pcpu_mem_free(chunk->map, size); 408 409 chunk->map_alloc = new_alloc; 410 chunk->map = new; 411 return 0; 412} 413 414/** 415 * pcpu_split_block - split a map block 416 * @chunk: chunk of interest 417 * @i: index of map block to split 418 * @head: head size in bytes (can be 0) 419 * @tail: tail size in bytes (can be 0) 420 * 421 * Split the @i'th map block into two or three blocks. If @head is 422 * non-zero, @head bytes block is inserted before block @i moving it 423 * to @i+1 and reducing its size by @head bytes. 424 * 425 * If @tail is non-zero, the target block, which can be @i or @i+1 426 * depending on @head, is reduced by @tail bytes and @tail byte block 427 * is inserted after the target block. 428 * 429 * @chunk->map must have enough free slots to accomodate the split. 430 * 431 * CONTEXT: 432 * pcpu_lock. 433 */ 434static void pcpu_split_block(struct pcpu_chunk *chunk, int i, 435 int head, int tail) 436{ 437 int nr_extra = !!head + !!tail; 438 439 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); 440 441 /* insert new subblocks */ 442 memmove(&chunk->map[i + nr_extra], &chunk->map[i], 443 sizeof(chunk->map[0]) * (chunk->map_used - i)); 444 chunk->map_used += nr_extra; 445 446 if (head) { 447 chunk->map[i + 1] = chunk->map[i] - head; 448 chunk->map[i++] = head; 449 } 450 if (tail) { 451 chunk->map[i++] -= tail; 452 chunk->map[i] = tail; 453 } 454} 455 456/** 457 * pcpu_alloc_area - allocate area from a pcpu_chunk 458 * @chunk: chunk of interest 459 * @size: wanted size in bytes 460 * @align: wanted align 461 * 462 * Try to allocate @size bytes area aligned at @align from @chunk. 463 * Note that this function only allocates the offset. It doesn't 464 * populate or map the area. 465 * 466 * @chunk->map must have at least two free slots. 467 * 468 * CONTEXT: 469 * pcpu_lock. 470 * 471 * RETURNS: 472 * Allocated offset in @chunk on success, -1 if no matching area is 473 * found. 474 */ 475static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) 476{ 477 int oslot = pcpu_chunk_slot(chunk); 478 int max_contig = 0; 479 int i, off; 480 481 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { 482 bool is_last = i + 1 == chunk->map_used; 483 int head, tail; 484 485 /* extra for alignment requirement */ 486 head = ALIGN(off, align) - off; 487 BUG_ON(i == 0 && head != 0); 488 489 if (chunk->map[i] < 0) 490 continue; 491 if (chunk->map[i] < head + size) { 492 max_contig = max(chunk->map[i], max_contig); 493 continue; 494 } 495 496 /* 497 * If head is small or the previous block is free, 498 * merge'em. Note that 'small' is defined as smaller 499 * than sizeof(int), which is very small but isn't too 500 * uncommon for percpu allocations. 501 */ 502 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { 503 if (chunk->map[i - 1] > 0) 504 chunk->map[i - 1] += head; 505 else { 506 chunk->map[i - 1] -= head; 507 chunk->free_size -= head; 508 } 509 chunk->map[i] -= head; 510 off += head; 511 head = 0; 512 } 513 514 /* if tail is small, just keep it around */ 515 tail = chunk->map[i] - head - size; 516 if (tail < sizeof(int)) 517 tail = 0; 518 519 /* split if warranted */ 520 if (head || tail) { 521 pcpu_split_block(chunk, i, head, tail); 522 if (head) { 523 i++; 524 off += head; 525 max_contig = max(chunk->map[i - 1], max_contig); 526 } 527 if (tail) 528 max_contig = max(chunk->map[i + 1], max_contig); 529 } 530 531 /* update hint and mark allocated */ 532 if (is_last) 533 chunk->contig_hint = max_contig; /* fully scanned */ 534 else 535 chunk->contig_hint = max(chunk->contig_hint, 536 max_contig); 537 538 chunk->free_size -= chunk->map[i]; 539 chunk->map[i] = -chunk->map[i]; 540 541 pcpu_chunk_relocate(chunk, oslot); 542 return off; 543 } 544 545 chunk->contig_hint = max_contig; /* fully scanned */ 546 pcpu_chunk_relocate(chunk, oslot); 547 548 /* tell the upper layer that this chunk has no matching area */ 549 return -1; 550} 551 552/** 553 * pcpu_free_area - free area to a pcpu_chunk 554 * @chunk: chunk of interest 555 * @freeme: offset of area to free 556 * 557 * Free area starting from @freeme to @chunk. Note that this function 558 * only modifies the allocation map. It doesn't depopulate or unmap 559 * the area. 560 * 561 * CONTEXT: 562 * pcpu_lock. 563 */ 564static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) 565{ 566 int oslot = pcpu_chunk_slot(chunk); 567 int i, off; 568 569 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) 570 if (off == freeme) 571 break; 572 BUG_ON(off != freeme); 573 BUG_ON(chunk->map[i] > 0); 574 575 chunk->map[i] = -chunk->map[i]; 576 chunk->free_size += chunk->map[i]; 577 578 /* merge with previous? */ 579 if (i > 0 && chunk->map[i - 1] >= 0) { 580 chunk->map[i - 1] += chunk->map[i]; 581 chunk->map_used--; 582 memmove(&chunk->map[i], &chunk->map[i + 1], 583 (chunk->map_used - i) * sizeof(chunk->map[0])); 584 i--; 585 } 586 /* merge with next? */ 587 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { 588 chunk->map[i] += chunk->map[i + 1]; 589 chunk->map_used--; 590 memmove(&chunk->map[i + 1], &chunk->map[i + 2], 591 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); 592 } 593 594 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); 595 pcpu_chunk_relocate(chunk, oslot); 596} 597 598/** 599 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap 600 * @chunk: chunk of interest 601 * @bitmapp: output parameter for bitmap 602 * @may_alloc: may allocate the array 603 * 604 * Returns pointer to array of pointers to struct page and bitmap, 605 * both of which can be indexed with pcpu_page_idx(). The returned 606 * array is cleared to zero and *@bitmapp is copied from 607 * @chunk->populated. Note that there is only one array and bitmap 608 * and access exclusion is the caller's responsibility. 609 * 610 * CONTEXT: 611 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc. 612 * Otherwise, don't care. 613 * 614 * RETURNS: 615 * Pointer to temp pages array on success, NULL on failure. 616 */ 617static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk, 618 unsigned long **bitmapp, 619 bool may_alloc) 620{ 621 static struct page **pages; 622 static unsigned long *bitmap; 623 size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]); 624 size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) * 625 sizeof(unsigned long); 626 627 if (!pages || !bitmap) { 628 if (may_alloc && !pages) 629 pages = pcpu_mem_alloc(pages_size); 630 if (may_alloc && !bitmap) 631 bitmap = pcpu_mem_alloc(bitmap_size); 632 if (!pages || !bitmap) 633 return NULL; 634 } 635 636 memset(pages, 0, pages_size); 637 bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages); 638 639 *bitmapp = bitmap; 640 return pages; 641} 642 643/** 644 * pcpu_free_pages - free pages which were allocated for @chunk 645 * @chunk: chunk pages were allocated for 646 * @pages: array of pages to be freed, indexed by pcpu_page_idx() 647 * @populated: populated bitmap 648 * @page_start: page index of the first page to be freed 649 * @page_end: page index of the last page to be freed + 1 650 * 651 * Free pages [@page_start and @page_end) in @pages for all units. 652 * The pages were allocated for @chunk. 653 */ 654static void pcpu_free_pages(struct pcpu_chunk *chunk, 655 struct page **pages, unsigned long *populated, 656 int page_start, int page_end) 657{ 658 unsigned int cpu; 659 int i; 660 661 for_each_possible_cpu(cpu) { 662 for (i = page_start; i < page_end; i++) { 663 struct page *page = pages[pcpu_page_idx(cpu, i)]; 664 665 if (page) 666 __free_page(page); 667 } 668 } 669} 670 671/** 672 * pcpu_alloc_pages - allocates pages for @chunk 673 * @chunk: target chunk 674 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx() 675 * @populated: populated bitmap 676 * @page_start: page index of the first page to be allocated 677 * @page_end: page index of the last page to be allocated + 1 678 * 679 * Allocate pages [@page_start,@page_end) into @pages for all units. 680 * The allocation is for @chunk. Percpu core doesn't care about the 681 * content of @pages and will pass it verbatim to pcpu_map_pages(). 682 */ 683static int pcpu_alloc_pages(struct pcpu_chunk *chunk, 684 struct page **pages, unsigned long *populated, 685 int page_start, int page_end) 686{ 687 const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; 688 unsigned int cpu; 689 int i; 690 691 for_each_possible_cpu(cpu) { 692 for (i = page_start; i < page_end; i++) { 693 struct page **pagep = &pages[pcpu_page_idx(cpu, i)]; 694 695 *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0); 696 if (!*pagep) { 697 pcpu_free_pages(chunk, pages, populated, 698 page_start, page_end); 699 return -ENOMEM; 700 } 701 } 702 } 703 return 0; 704} 705 706/** 707 * pcpu_pre_unmap_flush - flush cache prior to unmapping 708 * @chunk: chunk the regions to be flushed belongs to 709 * @page_start: page index of the first page to be flushed 710 * @page_end: page index of the last page to be flushed + 1 711 * 712 * Pages in [@page_start,@page_end) of @chunk are about to be 713 * unmapped. Flush cache. As each flushing trial can be very 714 * expensive, issue flush on the whole region at once rather than 715 * doing it for each cpu. This could be an overkill but is more 716 * scalable. 717 */ 718static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk, 719 int page_start, int page_end) 720{ 721 flush_cache_vunmap( 722 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), 723 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); 724} 725 726static void __pcpu_unmap_pages(unsigned long addr, int nr_pages) 727{ 728 unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT); 729} 730 731/** 732 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk 733 * @chunk: chunk of interest 734 * @pages: pages array which can be used to pass information to free 735 * @populated: populated bitmap 736 * @page_start: page index of the first page to unmap 737 * @page_end: page index of the last page to unmap + 1 738 * 739 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. 740 * Corresponding elements in @pages were cleared by the caller and can 741 * be used to carry information to pcpu_free_pages() which will be 742 * called after all unmaps are finished. The caller should call 743 * proper pre/post flush functions. 744 */ 745static void pcpu_unmap_pages(struct pcpu_chunk *chunk, 746 struct page **pages, unsigned long *populated, 747 int page_start, int page_end) 748{ 749 unsigned int cpu; 750 int i; 751 752 for_each_possible_cpu(cpu) { 753 for (i = page_start; i < page_end; i++) { 754 struct page *page; 755 756 page = pcpu_chunk_page(chunk, cpu, i); 757 WARN_ON(!page); 758 pages[pcpu_page_idx(cpu, i)] = page; 759 } 760 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start), 761 page_end - page_start); 762 } 763 764 for (i = page_start; i < page_end; i++) 765 __clear_bit(i, populated); 766} 767 768/** 769 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping 770 * @chunk: pcpu_chunk the regions to be flushed belong to 771 * @page_start: page index of the first page to be flushed 772 * @page_end: page index of the last page to be flushed + 1 773 * 774 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush 775 * TLB for the regions. This can be skipped if the area is to be 776 * returned to vmalloc as vmalloc will handle TLB flushing lazily. 777 * 778 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once 779 * for the whole region. 780 */ 781static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk, 782 int page_start, int page_end) 783{ 784 flush_tlb_kernel_range( 785 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), 786 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); 787} 788 789static int __pcpu_map_pages(unsigned long addr, struct page **pages, 790 int nr_pages) 791{ 792 return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT, 793 PAGE_KERNEL, pages); 794} 795 796/** 797 * pcpu_map_pages - map pages into a pcpu_chunk 798 * @chunk: chunk of interest 799 * @pages: pages array containing pages to be mapped 800 * @populated: populated bitmap 801 * @page_start: page index of the first page to map 802 * @page_end: page index of the last page to map + 1 803 * 804 * For each cpu, map pages [@page_start,@page_end) into @chunk. The 805 * caller is responsible for calling pcpu_post_map_flush() after all 806 * mappings are complete. 807 * 808 * This function is responsible for setting corresponding bits in 809 * @chunk->populated bitmap and whatever is necessary for reverse 810 * lookup (addr -> chunk). 811 */ 812static int pcpu_map_pages(struct pcpu_chunk *chunk, 813 struct page **pages, unsigned long *populated, 814 int page_start, int page_end) 815{ 816 unsigned int cpu, tcpu; 817 int i, err; 818 819 for_each_possible_cpu(cpu) { 820 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start), 821 &pages[pcpu_page_idx(cpu, page_start)], 822 page_end - page_start); 823 if (err < 0) 824 goto err; 825 } 826 827 /* mapping successful, link chunk and mark populated */ 828 for (i = page_start; i < page_end; i++) { 829 for_each_possible_cpu(cpu) 830 pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)], 831 chunk); 832 __set_bit(i, populated); 833 } 834 835 return 0; 836 837err: 838 for_each_possible_cpu(tcpu) { 839 if (tcpu == cpu) 840 break; 841 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start), 842 page_end - page_start); 843 } 844 return err; 845} 846 847/** 848 * pcpu_post_map_flush - flush cache after mapping 849 * @chunk: pcpu_chunk the regions to be flushed belong to 850 * @page_start: page index of the first page to be flushed 851 * @page_end: page index of the last page to be flushed + 1 852 * 853 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush 854 * cache. 855 * 856 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once 857 * for the whole region. 858 */ 859static void pcpu_post_map_flush(struct pcpu_chunk *chunk, 860 int page_start, int page_end) 861{ 862 flush_cache_vmap( 863 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), 864 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); 865} 866 867/** 868 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk 869 * @chunk: chunk to depopulate 870 * @off: offset to the area to depopulate 871 * @size: size of the area to depopulate in bytes 872 * @flush: whether to flush cache and tlb or not 873 * 874 * For each cpu, depopulate and unmap pages [@page_start,@page_end) 875 * from @chunk. If @flush is true, vcache is flushed before unmapping 876 * and tlb after. 877 * 878 * CONTEXT: 879 * pcpu_alloc_mutex. 880 */ 881static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size) 882{ 883 int page_start = PFN_DOWN(off); 884 int page_end = PFN_UP(off + size); 885 struct page **pages; 886 unsigned long *populated; 887 int rs, re; 888 889 /* quick path, check whether it's empty already */ 890 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { 891 if (rs == page_start && re == page_end) 892 return; 893 break; 894 } 895 896 /* immutable chunks can't be depopulated */ 897 WARN_ON(chunk->immutable); 898 899 /* 900 * If control reaches here, there must have been at least one 901 * successful population attempt so the temp pages array must 902 * be available now. 903 */ 904 pages = pcpu_get_pages_and_bitmap(chunk, &populated, false); 905 BUG_ON(!pages); 906 907 /* unmap and free */ 908 pcpu_pre_unmap_flush(chunk, page_start, page_end); 909 910 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) 911 pcpu_unmap_pages(chunk, pages, populated, rs, re); 912 913 /* no need to flush tlb, vmalloc will handle it lazily */ 914 915 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) 916 pcpu_free_pages(chunk, pages, populated, rs, re); 917 918 /* commit new bitmap */ 919 bitmap_copy(chunk->populated, populated, pcpu_unit_pages); 920} 921 922/** 923 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk 924 * @chunk: chunk of interest 925 * @off: offset to the area to populate 926 * @size: size of the area to populate in bytes 927 * 928 * For each cpu, populate and map pages [@page_start,@page_end) into 929 * @chunk. The area is cleared on return. 930 * 931 * CONTEXT: 932 * pcpu_alloc_mutex, does GFP_KERNEL allocation. 933 */ 934static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) 935{ 936 int page_start = PFN_DOWN(off); 937 int page_end = PFN_UP(off + size); 938 int free_end = page_start, unmap_end = page_start; 939 struct page **pages; 940 unsigned long *populated; 941 unsigned int cpu; 942 int rs, re, rc; 943 944 /* quick path, check whether all pages are already there */ 945 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) { 946 if (rs == page_start && re == page_end) 947 goto clear; 948 break; 949 } 950 951 /* need to allocate and map pages, this chunk can't be immutable */ 952 WARN_ON(chunk->immutable); 953 954 pages = pcpu_get_pages_and_bitmap(chunk, &populated, true); 955 if (!pages) 956 return -ENOMEM; 957 958 /* alloc and map */ 959 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { 960 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re); 961 if (rc) 962 goto err_free; 963 free_end = re; 964 } 965 966 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { 967 rc = pcpu_map_pages(chunk, pages, populated, rs, re); 968 if (rc) 969 goto err_unmap; 970 unmap_end = re; 971 } 972 pcpu_post_map_flush(chunk, page_start, page_end); 973 974 /* commit new bitmap */ 975 bitmap_copy(chunk->populated, populated, pcpu_unit_pages); 976clear: 977 for_each_possible_cpu(cpu) 978 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); 979 return 0; 980 981err_unmap: 982 pcpu_pre_unmap_flush(chunk, page_start, unmap_end); 983 pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end) 984 pcpu_unmap_pages(chunk, pages, populated, rs, re); 985 pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end); 986err_free: 987 pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end) 988 pcpu_free_pages(chunk, pages, populated, rs, re); 989 return rc; 990} 991 992static void free_pcpu_chunk(struct pcpu_chunk *chunk) 993{ 994 if (!chunk) 995 return; 996 if (chunk->vms) 997 pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups); 998 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); 999 kfree(chunk); 1000} 1001 1002static struct pcpu_chunk *alloc_pcpu_chunk(void) 1003{ 1004 struct pcpu_chunk *chunk; 1005 1006 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); 1007 if (!chunk) 1008 return NULL; 1009 1010 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); 1011 chunk->map_alloc = PCPU_DFL_MAP_ALLOC; 1012 chunk->map[chunk->map_used++] = pcpu_unit_size; 1013 1014 chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes, 1015 pcpu_nr_groups, pcpu_atom_size, 1016 GFP_KERNEL); 1017 if (!chunk->vms) { 1018 free_pcpu_chunk(chunk); 1019 return NULL; 1020 } 1021 1022 INIT_LIST_HEAD(&chunk->list); 1023 chunk->free_size = pcpu_unit_size; 1024 chunk->contig_hint = pcpu_unit_size; 1025 chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0]; 1026 1027 return chunk; 1028} 1029 1030/** 1031 * pcpu_alloc - the percpu allocator 1032 * @size: size of area to allocate in bytes 1033 * @align: alignment of area (max PAGE_SIZE) 1034 * @reserved: allocate from the reserved chunk if available 1035 * 1036 * Allocate percpu area of @size bytes aligned at @align. 1037 * 1038 * CONTEXT: 1039 * Does GFP_KERNEL allocation. 1040 * 1041 * RETURNS: 1042 * Percpu pointer to the allocated area on success, NULL on failure. 1043 */ 1044static void *pcpu_alloc(size_t size, size_t align, bool reserved) 1045{ 1046 struct pcpu_chunk *chunk; 1047 int slot, off; 1048 1049 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { 1050 WARN(true, "illegal size (%zu) or align (%zu) for " 1051 "percpu allocation\n", size, align); 1052 return NULL; 1053 } 1054 1055 mutex_lock(&pcpu_alloc_mutex); 1056 spin_lock_irq(&pcpu_lock); 1057 1058 /* serve reserved allocations from the reserved chunk if available */ 1059 if (reserved && pcpu_reserved_chunk) { 1060 chunk = pcpu_reserved_chunk; 1061 if (size > chunk->contig_hint || 1062 pcpu_extend_area_map(chunk) < 0) 1063 goto fail_unlock; 1064 off = pcpu_alloc_area(chunk, size, align); 1065 if (off >= 0) 1066 goto area_found; 1067 goto fail_unlock; 1068 } 1069 1070restart: 1071 /* search through normal chunks */ 1072 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 1073 list_for_each_entry(chunk, &pcpu_slot[slot], list) { 1074 if (size > chunk->contig_hint) 1075 continue; 1076 1077 switch (pcpu_extend_area_map(chunk)) { 1078 case 0: 1079 break; 1080 case 1: 1081 goto restart; /* pcpu_lock dropped, restart */ 1082 default: 1083 goto fail_unlock; 1084 } 1085 1086 off = pcpu_alloc_area(chunk, size, align); 1087 if (off >= 0) 1088 goto area_found; 1089 } 1090 } 1091 1092 /* hmmm... no space left, create a new chunk */ 1093 spin_unlock_irq(&pcpu_lock); 1094 1095 chunk = alloc_pcpu_chunk(); 1096 if (!chunk) 1097 goto fail_unlock_mutex; 1098 1099 spin_lock_irq(&pcpu_lock); 1100 pcpu_chunk_relocate(chunk, -1); 1101 goto restart; 1102 1103area_found: 1104 spin_unlock_irq(&pcpu_lock); 1105 1106 /* populate, map and clear the area */ 1107 if (pcpu_populate_chunk(chunk, off, size)) { 1108 spin_lock_irq(&pcpu_lock); 1109 pcpu_free_area(chunk, off); 1110 goto fail_unlock; 1111 } 1112 1113 mutex_unlock(&pcpu_alloc_mutex); 1114 1115 /* return address relative to base address */ 1116 return __addr_to_pcpu_ptr(chunk->base_addr + off); 1117 1118fail_unlock: 1119 spin_unlock_irq(&pcpu_lock); 1120fail_unlock_mutex: 1121 mutex_unlock(&pcpu_alloc_mutex); 1122 return NULL; 1123} 1124 1125/** 1126 * __alloc_percpu - allocate dynamic percpu area 1127 * @size: size of area to allocate in bytes 1128 * @align: alignment of area (max PAGE_SIZE) 1129 * 1130 * Allocate percpu area of @size bytes aligned at @align. Might 1131 * sleep. Might trigger writeouts. 1132 * 1133 * CONTEXT: 1134 * Does GFP_KERNEL allocation. 1135 * 1136 * RETURNS: 1137 * Percpu pointer to the allocated area on success, NULL on failure. 1138 */ 1139void *__alloc_percpu(size_t size, size_t align) 1140{ 1141 return pcpu_alloc(size, align, false); 1142} 1143EXPORT_SYMBOL_GPL(__alloc_percpu); 1144 1145/** 1146 * __alloc_reserved_percpu - allocate reserved percpu area 1147 * @size: size of area to allocate in bytes 1148 * @align: alignment of area (max PAGE_SIZE) 1149 * 1150 * Allocate percpu area of @size bytes aligned at @align from reserved 1151 * percpu area if arch has set it up; otherwise, allocation is served 1152 * from the same dynamic area. Might sleep. Might trigger writeouts. 1153 * 1154 * CONTEXT: 1155 * Does GFP_KERNEL allocation. 1156 * 1157 * RETURNS: 1158 * Percpu pointer to the allocated area on success, NULL on failure. 1159 */ 1160void *__alloc_reserved_percpu(size_t size, size_t align) 1161{ 1162 return pcpu_alloc(size, align, true); 1163} 1164 1165/** 1166 * pcpu_reclaim - reclaim fully free chunks, workqueue function 1167 * @work: unused 1168 * 1169 * Reclaim all fully free chunks except for the first one. 1170 * 1171 * CONTEXT: 1172 * workqueue context. 1173 */ 1174static void pcpu_reclaim(struct work_struct *work) 1175{ 1176 LIST_HEAD(todo); 1177 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; 1178 struct pcpu_chunk *chunk, *next; 1179 1180 mutex_lock(&pcpu_alloc_mutex); 1181 spin_lock_irq(&pcpu_lock); 1182 1183 list_for_each_entry_safe(chunk, next, head, list) { 1184 WARN_ON(chunk->immutable); 1185 1186 /* spare the first one */ 1187 if (chunk == list_first_entry(head, struct pcpu_chunk, list)) 1188 continue; 1189 1190 list_move(&chunk->list, &todo); 1191 } 1192 1193 spin_unlock_irq(&pcpu_lock); 1194 1195 list_for_each_entry_safe(chunk, next, &todo, list) { 1196 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size); 1197 free_pcpu_chunk(chunk); 1198 } 1199 1200 mutex_unlock(&pcpu_alloc_mutex); 1201} 1202 1203/** 1204 * free_percpu - free percpu area 1205 * @ptr: pointer to area to free 1206 * 1207 * Free percpu area @ptr. 1208 * 1209 * CONTEXT: 1210 * Can be called from atomic context. 1211 */ 1212void free_percpu(void *ptr) 1213{ 1214 void *addr = __pcpu_ptr_to_addr(ptr); 1215 struct pcpu_chunk *chunk; 1216 unsigned long flags; 1217 int off; 1218 1219 if (!ptr) 1220 return; 1221 1222 spin_lock_irqsave(&pcpu_lock, flags); 1223 1224 chunk = pcpu_chunk_addr_search(addr); 1225 off = addr - chunk->base_addr; 1226 1227 pcpu_free_area(chunk, off); 1228 1229 /* if there are more than one fully free chunks, wake up grim reaper */ 1230 if (chunk->free_size == pcpu_unit_size) { 1231 struct pcpu_chunk *pos; 1232 1233 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 1234 if (pos != chunk) { 1235 schedule_work(&pcpu_reclaim_work); 1236 break; 1237 } 1238 } 1239 1240 spin_unlock_irqrestore(&pcpu_lock, flags); 1241} 1242EXPORT_SYMBOL_GPL(free_percpu); 1243 1244static inline size_t pcpu_calc_fc_sizes(size_t static_size, 1245 size_t reserved_size, 1246 ssize_t *dyn_sizep) 1247{ 1248 size_t size_sum; 1249 1250 size_sum = PFN_ALIGN(static_size + reserved_size + 1251 (*dyn_sizep >= 0 ? *dyn_sizep : 0)); 1252 if (*dyn_sizep != 0) 1253 *dyn_sizep = size_sum - static_size - reserved_size; 1254 1255 return size_sum; 1256} 1257 1258/** 1259 * pcpu_alloc_alloc_info - allocate percpu allocation info 1260 * @nr_groups: the number of groups 1261 * @nr_units: the number of units 1262 * 1263 * Allocate ai which is large enough for @nr_groups groups containing 1264 * @nr_units units. The returned ai's groups[0].cpu_map points to the 1265 * cpu_map array which is long enough for @nr_units and filled with 1266 * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1267 * pointer of other groups. 1268 * 1269 * RETURNS: 1270 * Pointer to the allocated pcpu_alloc_info on success, NULL on 1271 * failure. 1272 */ 1273struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 1274 int nr_units) 1275{ 1276 struct pcpu_alloc_info *ai; 1277 size_t base_size, ai_size; 1278 void *ptr; 1279 int unit; 1280 1281 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 1282 __alignof__(ai->groups[0].cpu_map[0])); 1283 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 1284 1285 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size)); 1286 if (!ptr) 1287 return NULL; 1288 ai = ptr; 1289 ptr += base_size; 1290 1291 ai->groups[0].cpu_map = ptr; 1292 1293 for (unit = 0; unit < nr_units; unit++) 1294 ai->groups[0].cpu_map[unit] = NR_CPUS; 1295 1296 ai->nr_groups = nr_groups; 1297 ai->__ai_size = PFN_ALIGN(ai_size); 1298 1299 return ai; 1300} 1301 1302/** 1303 * pcpu_free_alloc_info - free percpu allocation info 1304 * @ai: pcpu_alloc_info to free 1305 * 1306 * Free @ai which was allocated by pcpu_alloc_alloc_info(). 1307 */ 1308void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 1309{ 1310 free_bootmem(__pa(ai), ai->__ai_size); 1311} 1312 1313/** 1314 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 1315 * @reserved_size: the size of reserved percpu area in bytes 1316 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto 1317 * @atom_size: allocation atom size 1318 * @cpu_distance_fn: callback to determine distance between cpus, optional 1319 * 1320 * This function determines grouping of units, their mappings to cpus 1321 * and other parameters considering needed percpu size, allocation 1322 * atom size and distances between CPUs. 1323 * 1324 * Groups are always mutliples of atom size and CPUs which are of 1325 * LOCAL_DISTANCE both ways are grouped together and share space for 1326 * units in the same group. The returned configuration is guaranteed 1327 * to have CPUs on different nodes on different groups and >=75% usage 1328 * of allocated virtual address space. 1329 * 1330 * RETURNS: 1331 * On success, pointer to the new allocation_info is returned. On 1332 * failure, ERR_PTR value is returned. 1333 */ 1334struct pcpu_alloc_info * __init pcpu_build_alloc_info( 1335 size_t reserved_size, ssize_t dyn_size, 1336 size_t atom_size, 1337 pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 1338{ 1339 static int group_map[NR_CPUS] __initdata; 1340 static int group_cnt[NR_CPUS] __initdata; 1341 const size_t static_size = __per_cpu_end - __per_cpu_start; 1342 int group_cnt_max = 0, nr_groups = 1, nr_units = 0; 1343 size_t size_sum, min_unit_size, alloc_size; 1344 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 1345 int last_allocs, group, unit; 1346 unsigned int cpu, tcpu; 1347 struct pcpu_alloc_info *ai; 1348 unsigned int *cpu_map; 1349 1350 /* 1351 * Determine min_unit_size, alloc_size and max_upa such that 1352 * alloc_size is multiple of atom_size and is the smallest 1353 * which can accomodate 4k aligned segments which are equal to 1354 * or larger than min_unit_size. 1355 */ 1356 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size); 1357 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 1358 1359 alloc_size = roundup(min_unit_size, atom_size); 1360 upa = alloc_size / min_unit_size; 1361 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1362 upa--; 1363 max_upa = upa; 1364 1365 /* group cpus according to their proximity */ 1366 for_each_possible_cpu(cpu) { 1367 group = 0; 1368 next_group: 1369 for_each_possible_cpu(tcpu) { 1370 if (cpu == tcpu) 1371 break; 1372 if (group_map[tcpu] == group && cpu_distance_fn && 1373 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 1374 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 1375 group++; 1376 nr_groups = max(nr_groups, group + 1); 1377 goto next_group; 1378 } 1379 } 1380 group_map[cpu] = group; 1381 group_cnt[group]++; 1382 group_cnt_max = max(group_cnt_max, group_cnt[group]); 1383 } 1384 1385 /* 1386 * Expand unit size until address space usage goes over 75% 1387 * and then as much as possible without using more address 1388 * space. 1389 */ 1390 last_allocs = INT_MAX; 1391 for (upa = max_upa; upa; upa--) { 1392 int allocs = 0, wasted = 0; 1393 1394 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1395 continue; 1396 1397 for (group = 0; group < nr_groups; group++) { 1398 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 1399 allocs += this_allocs; 1400 wasted += this_allocs * upa - group_cnt[group]; 1401 } 1402 1403 /* 1404 * Don't accept if wastage is over 25%. The 1405 * greater-than comparison ensures upa==1 always 1406 * passes the following check. 1407 */ 1408 if (wasted > num_possible_cpus() / 3) 1409 continue; 1410 1411 /* and then don't consume more memory */ 1412 if (allocs > last_allocs) 1413 break; 1414 last_allocs = allocs; 1415 best_upa = upa; 1416 } 1417 upa = best_upa; 1418 1419 /* allocate and fill alloc_info */ 1420 for (group = 0; group < nr_groups; group++) 1421 nr_units += roundup(group_cnt[group], upa); 1422 1423 ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 1424 if (!ai) 1425 return ERR_PTR(-ENOMEM); 1426 cpu_map = ai->groups[0].cpu_map; 1427 1428 for (group = 0; group < nr_groups; group++) { 1429 ai->groups[group].cpu_map = cpu_map; 1430 cpu_map += roundup(group_cnt[group], upa); 1431 } 1432 1433 ai->static_size = static_size; 1434 ai->reserved_size = reserved_size; 1435 ai->dyn_size = dyn_size; 1436 ai->unit_size = alloc_size / upa; 1437 ai->atom_size = atom_size; 1438 ai->alloc_size = alloc_size; 1439 1440 for (group = 0, unit = 0; group_cnt[group]; group++) { 1441 struct pcpu_group_info *gi = &ai->groups[group]; 1442 1443 /* 1444 * Initialize base_offset as if all groups are located 1445 * back-to-back. The caller should update this to 1446 * reflect actual allocation. 1447 */ 1448 gi->base_offset = unit * ai->unit_size; 1449 1450 for_each_possible_cpu(cpu) 1451 if (group_map[cpu] == group) 1452 gi->cpu_map[gi->nr_units++] = cpu; 1453 gi->nr_units = roundup(gi->nr_units, upa); 1454 unit += gi->nr_units; 1455 } 1456 BUG_ON(unit != nr_units); 1457 1458 return ai; 1459} 1460 1461/** 1462 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 1463 * @lvl: loglevel 1464 * @ai: allocation info to dump 1465 * 1466 * Print out information about @ai using loglevel @lvl. 1467 */ 1468static void pcpu_dump_alloc_info(const char *lvl, 1469 const struct pcpu_alloc_info *ai) 1470{ 1471 int group_width = 1, cpu_width = 1, width; 1472 char empty_str[] = "--------"; 1473 int alloc = 0, alloc_end = 0; 1474 int group, v; 1475 int upa, apl; /* units per alloc, allocs per line */ 1476 1477 v = ai->nr_groups; 1478 while (v /= 10) 1479 group_width++; 1480 1481 v = num_possible_cpus(); 1482 while (v /= 10) 1483 cpu_width++; 1484 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 1485 1486 upa = ai->alloc_size / ai->unit_size; 1487 width = upa * (cpu_width + 1) + group_width + 3; 1488 apl = rounddown_pow_of_two(max(60 / width, 1)); 1489 1490 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 1491 lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 1492 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 1493 1494 for (group = 0; group < ai->nr_groups; group++) { 1495 const struct pcpu_group_info *gi = &ai->groups[group]; 1496 int unit = 0, unit_end = 0; 1497 1498 BUG_ON(gi->nr_units % upa); 1499 for (alloc_end += gi->nr_units / upa; 1500 alloc < alloc_end; alloc++) { 1501 if (!(alloc % apl)) { 1502 printk("\n"); 1503 printk("%spcpu-alloc: ", lvl); 1504 } 1505 printk("[%0*d] ", group_width, group); 1506 1507 for (unit_end += upa; unit < unit_end; unit++) 1508 if (gi->cpu_map[unit] != NR_CPUS) 1509 printk("%0*d ", cpu_width, 1510 gi->cpu_map[unit]); 1511 else 1512 printk("%s ", empty_str); 1513 } 1514 } 1515 printk("\n"); 1516} 1517 1518/** 1519 * pcpu_setup_first_chunk - initialize the first percpu chunk 1520 * @ai: pcpu_alloc_info describing how to percpu area is shaped 1521 * @base_addr: mapped address 1522 * 1523 * Initialize the first percpu chunk which contains the kernel static 1524 * perpcu area. This function is to be called from arch percpu area 1525 * setup path. 1526 * 1527 * @ai contains all information necessary to initialize the first 1528 * chunk and prime the dynamic percpu allocator. 1529 * 1530 * @ai->static_size is the size of static percpu area. 1531 * 1532 * @ai->reserved_size, if non-zero, specifies the amount of bytes to 1533 * reserve after the static area in the first chunk. This reserves 1534 * the first chunk such that it's available only through reserved 1535 * percpu allocation. This is primarily used to serve module percpu 1536 * static areas on architectures where the addressing model has 1537 * limited offset range for symbol relocations to guarantee module 1538 * percpu symbols fall inside the relocatable range. 1539 * 1540 * @ai->dyn_size determines the number of bytes available for dynamic 1541 * allocation in the first chunk. The area between @ai->static_size + 1542 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 1543 * 1544 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 1545 * and equal to or larger than @ai->static_size + @ai->reserved_size + 1546 * @ai->dyn_size. 1547 * 1548 * @ai->atom_size is the allocation atom size and used as alignment 1549 * for vm areas. 1550 * 1551 * @ai->alloc_size is the allocation size and always multiple of 1552 * @ai->atom_size. This is larger than @ai->atom_size if 1553 * @ai->unit_size is larger than @ai->atom_size. 1554 * 1555 * @ai->nr_groups and @ai->groups describe virtual memory layout of 1556 * percpu areas. Units which should be colocated are put into the 1557 * same group. Dynamic VM areas will be allocated according to these 1558 * groupings. If @ai->nr_groups is zero, a single group containing 1559 * all units is assumed. 1560 * 1561 * The caller should have mapped the first chunk at @base_addr and 1562 * copied static data to each unit. 1563 * 1564 * If the first chunk ends up with both reserved and dynamic areas, it 1565 * is served by two chunks - one to serve the core static and reserved 1566 * areas and the other for the dynamic area. They share the same vm 1567 * and page map but uses different area allocation map to stay away 1568 * from each other. The latter chunk is circulated in the chunk slots 1569 * and available for dynamic allocation like any other chunks. 1570 * 1571 * RETURNS: 1572 * 0 on success, -errno on failure. 1573 */ 1574int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 1575 void *base_addr) 1576{ 1577 static int smap[2], dmap[2]; 1578 size_t dyn_size = ai->dyn_size; 1579 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; 1580 struct pcpu_chunk *schunk, *dchunk = NULL; 1581 unsigned long *group_offsets; 1582 size_t *group_sizes; 1583 unsigned long *unit_off; 1584 unsigned int cpu; 1585 int *unit_map; 1586 int group, unit, i; 1587 1588 /* sanity checks */ 1589 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || 1590 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); 1591 BUG_ON(ai->nr_groups <= 0); 1592 BUG_ON(!ai->static_size); 1593 BUG_ON(!base_addr); 1594 BUG_ON(ai->unit_size < size_sum); 1595 BUG_ON(ai->unit_size & ~PAGE_MASK); 1596 BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 1597 1598 pcpu_dump_alloc_info(KERN_DEBUG, ai); 1599 1600 /* process group information and build config tables accordingly */ 1601 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0])); 1602 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0])); 1603 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0])); 1604 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0])); 1605 1606 for (cpu = 0; cpu < nr_cpu_ids; cpu++) 1607 unit_map[cpu] = NR_CPUS; 1608 pcpu_first_unit_cpu = NR_CPUS; 1609 1610 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 1611 const struct pcpu_group_info *gi = &ai->groups[group]; 1612 1613 group_offsets[group] = gi->base_offset; 1614 group_sizes[group] = gi->nr_units * ai->unit_size; 1615 1616 for (i = 0; i < gi->nr_units; i++) { 1617 cpu = gi->cpu_map[i]; 1618 if (cpu == NR_CPUS) 1619 continue; 1620 1621 BUG_ON(cpu > nr_cpu_ids || !cpu_possible(cpu)); 1622 BUG_ON(unit_map[cpu] != NR_CPUS); 1623 1624 unit_map[cpu] = unit + i; 1625 unit_off[cpu] = gi->base_offset + i * ai->unit_size; 1626 1627 if (pcpu_first_unit_cpu == NR_CPUS) 1628 pcpu_first_unit_cpu = cpu; 1629 } 1630 } 1631 pcpu_last_unit_cpu = cpu; 1632 pcpu_nr_units = unit; 1633 1634 for_each_possible_cpu(cpu) 1635 BUG_ON(unit_map[cpu] == NR_CPUS); 1636 1637 pcpu_nr_groups = ai->nr_groups; 1638 pcpu_group_offsets = group_offsets; 1639 pcpu_group_sizes = group_sizes; 1640 pcpu_unit_map = unit_map; 1641 pcpu_unit_offsets = unit_off; 1642 1643 /* determine basic parameters */ 1644 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 1645 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 1646 pcpu_atom_size = ai->atom_size; 1647 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 1648 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 1649 1650 /* 1651 * Allocate chunk slots. The additional last slot is for 1652 * empty chunks. 1653 */ 1654 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 1655 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); 1656 for (i = 0; i < pcpu_nr_slots; i++) 1657 INIT_LIST_HEAD(&pcpu_slot[i]); 1658 1659 /* 1660 * Initialize static chunk. If reserved_size is zero, the 1661 * static chunk covers static area + dynamic allocation area 1662 * in the first chunk. If reserved_size is not zero, it 1663 * covers static area + reserved area (mostly used for module 1664 * static percpu allocation). 1665 */ 1666 schunk = alloc_bootmem(pcpu_chunk_struct_size); 1667 INIT_LIST_HEAD(&schunk->list); 1668 schunk->base_addr = base_addr; 1669 schunk->map = smap; 1670 schunk->map_alloc = ARRAY_SIZE(smap); 1671 schunk->immutable = true; 1672 bitmap_fill(schunk->populated, pcpu_unit_pages); 1673 1674 if (ai->reserved_size) { 1675 schunk->free_size = ai->reserved_size; 1676 pcpu_reserved_chunk = schunk; 1677 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; 1678 } else { 1679 schunk->free_size = dyn_size; 1680 dyn_size = 0; /* dynamic area covered */ 1681 } 1682 schunk->contig_hint = schunk->free_size; 1683 1684 schunk->map[schunk->map_used++] = -ai->static_size; 1685 if (schunk->free_size) 1686 schunk->map[schunk->map_used++] = schunk->free_size; 1687 1688 /* init dynamic chunk if necessary */ 1689 if (dyn_size) { 1690 dchunk = alloc_bootmem(pcpu_chunk_struct_size); 1691 INIT_LIST_HEAD(&dchunk->list); 1692 dchunk->base_addr = base_addr; 1693 dchunk->map = dmap; 1694 dchunk->map_alloc = ARRAY_SIZE(dmap); 1695 dchunk->immutable = true; 1696 bitmap_fill(dchunk->populated, pcpu_unit_pages); 1697 1698 dchunk->contig_hint = dchunk->free_size = dyn_size; 1699 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; 1700 dchunk->map[dchunk->map_used++] = dchunk->free_size; 1701 } 1702 1703 /* link the first chunk in */ 1704 pcpu_first_chunk = dchunk ?: schunk; 1705 pcpu_chunk_relocate(pcpu_first_chunk, -1); 1706 1707 /* we're done */ 1708 pcpu_base_addr = base_addr; 1709 return 0; 1710} 1711 1712const char *pcpu_fc_names[PCPU_FC_NR] __initdata = { 1713 [PCPU_FC_AUTO] = "auto", 1714 [PCPU_FC_EMBED] = "embed", 1715 [PCPU_FC_PAGE] = "page", 1716}; 1717 1718enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 1719 1720static int __init percpu_alloc_setup(char *str) 1721{ 1722 if (0) 1723 /* nada */; 1724#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 1725 else if (!strcmp(str, "embed")) 1726 pcpu_chosen_fc = PCPU_FC_EMBED; 1727#endif 1728#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1729 else if (!strcmp(str, "page")) 1730 pcpu_chosen_fc = PCPU_FC_PAGE; 1731#endif 1732 else 1733 pr_warning("PERCPU: unknown allocator %s specified\n", str); 1734 1735 return 0; 1736} 1737early_param("percpu_alloc", percpu_alloc_setup); 1738 1739#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 1740 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 1741/** 1742 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 1743 * @reserved_size: the size of reserved percpu area in bytes 1744 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto 1745 * @atom_size: allocation atom size 1746 * @cpu_distance_fn: callback to determine distance between cpus, optional 1747 * @alloc_fn: function to allocate percpu page 1748 * @free_fn: funtion to free percpu page 1749 * 1750 * This is a helper to ease setting up embedded first percpu chunk and 1751 * can be called where pcpu_setup_first_chunk() is expected. 1752 * 1753 * If this function is used to setup the first chunk, it is allocated 1754 * by calling @alloc_fn and used as-is without being mapped into 1755 * vmalloc area. Allocations are always whole multiples of @atom_size 1756 * aligned to @atom_size. 1757 * 1758 * This enables the first chunk to piggy back on the linear physical 1759 * mapping which often uses larger page size. Please note that this 1760 * can result in very sparse cpu->unit mapping on NUMA machines thus 1761 * requiring large vmalloc address space. Don't use this allocator if 1762 * vmalloc space is not orders of magnitude larger than distances 1763 * between node memory addresses (ie. 32bit NUMA machines). 1764 * 1765 * When @dyn_size is positive, dynamic area might be larger than 1766 * specified to fill page alignment. When @dyn_size is auto, 1767 * @dyn_size is just big enough to fill page alignment after static 1768 * and reserved areas. 1769 * 1770 * If the needed size is smaller than the minimum or specified unit 1771 * size, the leftover is returned using @free_fn. 1772 * 1773 * RETURNS: 1774 * 0 on success, -errno on failure. 1775 */ 1776int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size, 1777 size_t atom_size, 1778 pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 1779 pcpu_fc_alloc_fn_t alloc_fn, 1780 pcpu_fc_free_fn_t free_fn) 1781{ 1782 void *base = (void *)ULONG_MAX; 1783 void **areas = NULL; 1784 struct pcpu_alloc_info *ai; 1785 size_t size_sum, areas_size; 1786 int group, i, rc; 1787 1788 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 1789 cpu_distance_fn); 1790 if (IS_ERR(ai)) 1791 return PTR_ERR(ai); 1792 1793 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 1794 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 1795 1796 areas = alloc_bootmem_nopanic(areas_size); 1797 if (!areas) { 1798 rc = -ENOMEM; 1799 goto out_free; 1800 } 1801 1802 /* allocate, copy and determine base address */ 1803 for (group = 0; group < ai->nr_groups; group++) { 1804 struct pcpu_group_info *gi = &ai->groups[group]; 1805 unsigned int cpu = NR_CPUS; 1806 void *ptr; 1807 1808 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 1809 cpu = gi->cpu_map[i]; 1810 BUG_ON(cpu == NR_CPUS); 1811 1812 /* allocate space for the whole group */ 1813 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 1814 if (!ptr) { 1815 rc = -ENOMEM; 1816 goto out_free_areas; 1817 } 1818 areas[group] = ptr; 1819 1820 base = min(ptr, base); 1821 1822 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 1823 if (gi->cpu_map[i] == NR_CPUS) { 1824 /* unused unit, free whole */ 1825 free_fn(ptr, ai->unit_size); 1826 continue; 1827 } 1828 /* copy and return the unused part */ 1829 memcpy(ptr, __per_cpu_load, ai->static_size); 1830 free_fn(ptr + size_sum, ai->unit_size - size_sum); 1831 } 1832 } 1833 1834 /* base address is now known, determine group base offsets */ 1835 for (group = 0; group < ai->nr_groups; group++) 1836 ai->groups[group].base_offset = areas[group] - base; 1837 1838 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 1839 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 1840 ai->dyn_size, ai->unit_size); 1841 1842 rc = pcpu_setup_first_chunk(ai, base); 1843 goto out_free; 1844 1845out_free_areas: 1846 for (group = 0; group < ai->nr_groups; group++) 1847 free_fn(areas[group], 1848 ai->groups[group].nr_units * ai->unit_size); 1849out_free: 1850 pcpu_free_alloc_info(ai); 1851 if (areas) 1852 free_bootmem(__pa(areas), areas_size); 1853 return rc; 1854} 1855#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK || 1856 !CONFIG_HAVE_SETUP_PER_CPU_AREA */ 1857 1858#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1859/** 1860 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 1861 * @reserved_size: the size of reserved percpu area in bytes 1862 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 1863 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE 1864 * @populate_pte_fn: function to populate pte 1865 * 1866 * This is a helper to ease setting up page-remapped first percpu 1867 * chunk and can be called where pcpu_setup_first_chunk() is expected. 1868 * 1869 * This is the basic allocator. Static percpu area is allocated 1870 * page-by-page into vmalloc area. 1871 * 1872 * RETURNS: 1873 * 0 on success, -errno on failure. 1874 */ 1875int __init pcpu_page_first_chunk(size_t reserved_size, 1876 pcpu_fc_alloc_fn_t alloc_fn, 1877 pcpu_fc_free_fn_t free_fn, 1878 pcpu_fc_populate_pte_fn_t populate_pte_fn) 1879{ 1880 static struct vm_struct vm; 1881 struct pcpu_alloc_info *ai; 1882 char psize_str[16]; 1883 int unit_pages; 1884 size_t pages_size; 1885 struct page **pages; 1886 int unit, i, j, rc; 1887 1888 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 1889 1890 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL); 1891 if (IS_ERR(ai)) 1892 return PTR_ERR(ai); 1893 BUG_ON(ai->nr_groups != 1); 1894 BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); 1895 1896 unit_pages = ai->unit_size >> PAGE_SHIFT; 1897 1898 /* unaligned allocations can't be freed, round up to page size */ 1899 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 1900 sizeof(pages[0])); 1901 pages = alloc_bootmem(pages_size); 1902 1903 /* allocate pages */ 1904 j = 0; 1905 for (unit = 0; unit < num_possible_cpus(); unit++) 1906 for (i = 0; i < unit_pages; i++) { 1907 unsigned int cpu = ai->groups[0].cpu_map[unit]; 1908 void *ptr; 1909 1910 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 1911 if (!ptr) { 1912 pr_warning("PERCPU: failed to allocate %s page " 1913 "for cpu%u\n", psize_str, cpu); 1914 goto enomem; 1915 } 1916 pages[j++] = virt_to_page(ptr); 1917 } 1918 1919 /* allocate vm area, map the pages and copy static data */ 1920 vm.flags = VM_ALLOC; 1921 vm.size = num_possible_cpus() * ai->unit_size; 1922 vm_area_register_early(&vm, PAGE_SIZE); 1923 1924 for (unit = 0; unit < num_possible_cpus(); unit++) { 1925 unsigned long unit_addr = 1926 (unsigned long)vm.addr + unit * ai->unit_size; 1927 1928 for (i = 0; i < unit_pages; i++) 1929 populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 1930 1931 /* pte already populated, the following shouldn't fail */ 1932 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 1933 unit_pages); 1934 if (rc < 0) 1935 panic("failed to map percpu area, err=%d\n", rc); 1936 1937 /* 1938 * FIXME: Archs with virtual cache should flush local 1939 * cache for the linear mapping here - something 1940 * equivalent to flush_cache_vmap() on the local cpu. 1941 * flush_cache_vmap() can't be used as most supporting 1942 * data structures are not set up yet. 1943 */ 1944 1945 /* copy static data */ 1946 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 1947 } 1948 1949 /* we're ready, commit */ 1950 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n", 1951 unit_pages, psize_str, vm.addr, ai->static_size, 1952 ai->reserved_size, ai->dyn_size); 1953 1954 rc = pcpu_setup_first_chunk(ai, vm.addr); 1955 goto out_free_ar; 1956 1957enomem: 1958 while (--j >= 0) 1959 free_fn(page_address(pages[j]), PAGE_SIZE); 1960 rc = -ENOMEM; 1961out_free_ar: 1962 free_bootmem(__pa(pages), pages_size); 1963 pcpu_free_alloc_info(ai); 1964 return rc; 1965} 1966#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */ 1967 1968/* 1969 * Generic percpu area setup. 1970 * 1971 * The embedding helper is used because its behavior closely resembles 1972 * the original non-dynamic generic percpu area setup. This is 1973 * important because many archs have addressing restrictions and might 1974 * fail if the percpu area is located far away from the previous 1975 * location. As an added bonus, in non-NUMA cases, embedding is 1976 * generally a good idea TLB-wise because percpu area can piggy back 1977 * on the physical linear memory mapping which uses large page 1978 * mappings on applicable archs. 1979 */ 1980#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 1981unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 1982EXPORT_SYMBOL(__per_cpu_offset); 1983 1984static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 1985 size_t align) 1986{ 1987 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS)); 1988} 1989 1990static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 1991{ 1992 free_bootmem(__pa(ptr), size); 1993} 1994 1995void __init setup_per_cpu_areas(void) 1996{ 1997 unsigned long delta; 1998 unsigned int cpu; 1999 int rc; 2000 2001 /* 2002 * Always reserve area for module percpu variables. That's 2003 * what the legacy allocator did. 2004 */ 2005 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 2006 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 2007 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 2008 if (rc < 0) 2009 panic("Failed to initialized percpu areas."); 2010 2011 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 2012 for_each_possible_cpu(cpu) 2013 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 2014} 2015#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 2016