percpu.c revision fb435d5233f8b6f9b93c11d6304d8e98fed03234
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 struct vm_struct *vm; /* mapped vmalloc region */ 98 int map_used; /* # of map entries used */ 99 int map_alloc; /* # of map entries allocated */ 100 int *map; /* allocation map */ 101 bool immutable; /* no [de]population allowed */ 102 unsigned long populated[]; /* populated bitmap */ 103}; 104 105static int pcpu_unit_pages __read_mostly; 106static int pcpu_unit_size __read_mostly; 107static int pcpu_nr_units __read_mostly; 108static int pcpu_chunk_size __read_mostly; 109static int pcpu_nr_slots __read_mostly; 110static size_t pcpu_chunk_struct_size __read_mostly; 111 112/* cpus with the lowest and highest unit numbers */ 113static unsigned int pcpu_first_unit_cpu __read_mostly; 114static unsigned int pcpu_last_unit_cpu __read_mostly; 115 116/* the address of the first chunk which starts with the kernel static area */ 117void *pcpu_base_addr __read_mostly; 118EXPORT_SYMBOL_GPL(pcpu_base_addr); 119 120static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */ 121const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */ 122 123/* 124 * The first chunk which always exists. Note that unlike other 125 * chunks, this one can be allocated and mapped in several different 126 * ways and thus often doesn't live in the vmalloc area. 127 */ 128static struct pcpu_chunk *pcpu_first_chunk; 129 130/* 131 * Optional reserved chunk. This chunk reserves part of the first 132 * chunk and serves it for reserved allocations. The amount of 133 * reserved offset is in pcpu_reserved_chunk_limit. When reserved 134 * area doesn't exist, the following variables contain NULL and 0 135 * respectively. 136 */ 137static struct pcpu_chunk *pcpu_reserved_chunk; 138static int pcpu_reserved_chunk_limit; 139 140/* 141 * Synchronization rules. 142 * 143 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former 144 * protects allocation/reclaim paths, chunks, populated bitmap and 145 * vmalloc mapping. The latter is a spinlock and protects the index 146 * data structures - chunk slots, chunks and area maps in chunks. 147 * 148 * During allocation, pcpu_alloc_mutex is kept locked all the time and 149 * pcpu_lock is grabbed and released as necessary. All actual memory 150 * allocations are done using GFP_KERNEL with pcpu_lock released. 151 * 152 * Free path accesses and alters only the index data structures, so it 153 * can be safely called from atomic context. When memory needs to be 154 * returned to the system, free path schedules reclaim_work which 155 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be 156 * reclaimed, release both locks and frees the chunks. Note that it's 157 * necessary to grab both locks to remove a chunk from circulation as 158 * allocation path might be referencing the chunk with only 159 * pcpu_alloc_mutex locked. 160 */ 161static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ 162static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ 163 164static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ 165 166/* reclaim work to release fully free chunks, scheduled from free path */ 167static void pcpu_reclaim(struct work_struct *work); 168static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); 169 170static int __pcpu_size_to_slot(int size) 171{ 172 int highbit = fls(size); /* size is in bytes */ 173 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 174} 175 176static int pcpu_size_to_slot(int size) 177{ 178 if (size == pcpu_unit_size) 179 return pcpu_nr_slots - 1; 180 return __pcpu_size_to_slot(size); 181} 182 183static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 184{ 185 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) 186 return 0; 187 188 return pcpu_size_to_slot(chunk->free_size); 189} 190 191static int pcpu_page_idx(unsigned int cpu, int page_idx) 192{ 193 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 194} 195 196static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 197 unsigned int cpu, int page_idx) 198{ 199 return (unsigned long)chunk->vm->addr + pcpu_unit_offsets[cpu] + 200 (page_idx << PAGE_SHIFT); 201} 202 203static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk, 204 unsigned int cpu, int page_idx) 205{ 206 /* must not be used on pre-mapped chunk */ 207 WARN_ON(chunk->immutable); 208 209 return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx)); 210} 211 212/* set the pointer to a chunk in a page struct */ 213static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 214{ 215 page->index = (unsigned long)pcpu; 216} 217 218/* obtain pointer to a chunk from a page struct */ 219static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 220{ 221 return (struct pcpu_chunk *)page->index; 222} 223 224static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end) 225{ 226 *rs = find_next_zero_bit(chunk->populated, end, *rs); 227 *re = find_next_bit(chunk->populated, end, *rs + 1); 228} 229 230static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end) 231{ 232 *rs = find_next_bit(chunk->populated, end, *rs); 233 *re = find_next_zero_bit(chunk->populated, end, *rs + 1); 234} 235 236/* 237 * (Un)populated page region iterators. Iterate over (un)populated 238 * page regions betwen @start and @end in @chunk. @rs and @re should 239 * be integer variables and will be set to start and end page index of 240 * the current region. 241 */ 242#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ 243 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ 244 (rs) < (re); \ 245 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) 246 247#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ 248 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ 249 (rs) < (re); \ 250 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) 251 252/** 253 * pcpu_mem_alloc - allocate memory 254 * @size: bytes to allocate 255 * 256 * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 257 * kzalloc() is used; otherwise, vmalloc() is used. The returned 258 * memory is always zeroed. 259 * 260 * CONTEXT: 261 * Does GFP_KERNEL allocation. 262 * 263 * RETURNS: 264 * Pointer to the allocated area on success, NULL on failure. 265 */ 266static void *pcpu_mem_alloc(size_t size) 267{ 268 if (size <= PAGE_SIZE) 269 return kzalloc(size, GFP_KERNEL); 270 else { 271 void *ptr = vmalloc(size); 272 if (ptr) 273 memset(ptr, 0, size); 274 return ptr; 275 } 276} 277 278/** 279 * pcpu_mem_free - free memory 280 * @ptr: memory to free 281 * @size: size of the area 282 * 283 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). 284 */ 285static void pcpu_mem_free(void *ptr, size_t size) 286{ 287 if (size <= PAGE_SIZE) 288 kfree(ptr); 289 else 290 vfree(ptr); 291} 292 293/** 294 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 295 * @chunk: chunk of interest 296 * @oslot: the previous slot it was on 297 * 298 * This function is called after an allocation or free changed @chunk. 299 * New slot according to the changed state is determined and @chunk is 300 * moved to the slot. Note that the reserved chunk is never put on 301 * chunk slots. 302 * 303 * CONTEXT: 304 * pcpu_lock. 305 */ 306static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 307{ 308 int nslot = pcpu_chunk_slot(chunk); 309 310 if (chunk != pcpu_reserved_chunk && oslot != nslot) { 311 if (oslot < nslot) 312 list_move(&chunk->list, &pcpu_slot[nslot]); 313 else 314 list_move_tail(&chunk->list, &pcpu_slot[nslot]); 315 } 316} 317 318/** 319 * pcpu_chunk_addr_search - determine chunk containing specified address 320 * @addr: address for which the chunk needs to be determined. 321 * 322 * RETURNS: 323 * The address of the found chunk. 324 */ 325static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 326{ 327 void *first_start = pcpu_first_chunk->vm->addr; 328 329 /* is it in the first chunk? */ 330 if (addr >= first_start && addr < first_start + pcpu_unit_size) { 331 /* is it in the reserved area? */ 332 if (addr < first_start + pcpu_reserved_chunk_limit) 333 return pcpu_reserved_chunk; 334 return pcpu_first_chunk; 335 } 336 337 /* 338 * The address is relative to unit0 which might be unused and 339 * thus unmapped. Offset the address to the unit space of the 340 * current processor before looking it up in the vmalloc 341 * space. Note that any possible cpu id can be used here, so 342 * there's no need to worry about preemption or cpu hotplug. 343 */ 344 addr += pcpu_unit_offsets[smp_processor_id()]; 345 return pcpu_get_page_chunk(vmalloc_to_page(addr)); 346} 347 348/** 349 * pcpu_extend_area_map - extend area map for allocation 350 * @chunk: target chunk 351 * 352 * Extend area map of @chunk so that it can accomodate an allocation. 353 * A single allocation can split an area into three areas, so this 354 * function makes sure that @chunk->map has at least two extra slots. 355 * 356 * CONTEXT: 357 * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired 358 * if area map is extended. 359 * 360 * RETURNS: 361 * 0 if noop, 1 if successfully extended, -errno on failure. 362 */ 363static int pcpu_extend_area_map(struct pcpu_chunk *chunk) 364{ 365 int new_alloc; 366 int *new; 367 size_t size; 368 369 /* has enough? */ 370 if (chunk->map_alloc >= chunk->map_used + 2) 371 return 0; 372 373 spin_unlock_irq(&pcpu_lock); 374 375 new_alloc = PCPU_DFL_MAP_ALLOC; 376 while (new_alloc < chunk->map_used + 2) 377 new_alloc *= 2; 378 379 new = pcpu_mem_alloc(new_alloc * sizeof(new[0])); 380 if (!new) { 381 spin_lock_irq(&pcpu_lock); 382 return -ENOMEM; 383 } 384 385 /* 386 * Acquire pcpu_lock and switch to new area map. Only free 387 * could have happened inbetween, so map_used couldn't have 388 * grown. 389 */ 390 spin_lock_irq(&pcpu_lock); 391 BUG_ON(new_alloc < chunk->map_used + 2); 392 393 size = chunk->map_alloc * sizeof(chunk->map[0]); 394 memcpy(new, chunk->map, size); 395 396 /* 397 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is 398 * one of the first chunks and still using static map. 399 */ 400 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC) 401 pcpu_mem_free(chunk->map, size); 402 403 chunk->map_alloc = new_alloc; 404 chunk->map = new; 405 return 0; 406} 407 408/** 409 * pcpu_split_block - split a map block 410 * @chunk: chunk of interest 411 * @i: index of map block to split 412 * @head: head size in bytes (can be 0) 413 * @tail: tail size in bytes (can be 0) 414 * 415 * Split the @i'th map block into two or three blocks. If @head is 416 * non-zero, @head bytes block is inserted before block @i moving it 417 * to @i+1 and reducing its size by @head bytes. 418 * 419 * If @tail is non-zero, the target block, which can be @i or @i+1 420 * depending on @head, is reduced by @tail bytes and @tail byte block 421 * is inserted after the target block. 422 * 423 * @chunk->map must have enough free slots to accomodate the split. 424 * 425 * CONTEXT: 426 * pcpu_lock. 427 */ 428static void pcpu_split_block(struct pcpu_chunk *chunk, int i, 429 int head, int tail) 430{ 431 int nr_extra = !!head + !!tail; 432 433 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); 434 435 /* insert new subblocks */ 436 memmove(&chunk->map[i + nr_extra], &chunk->map[i], 437 sizeof(chunk->map[0]) * (chunk->map_used - i)); 438 chunk->map_used += nr_extra; 439 440 if (head) { 441 chunk->map[i + 1] = chunk->map[i] - head; 442 chunk->map[i++] = head; 443 } 444 if (tail) { 445 chunk->map[i++] -= tail; 446 chunk->map[i] = tail; 447 } 448} 449 450/** 451 * pcpu_alloc_area - allocate area from a pcpu_chunk 452 * @chunk: chunk of interest 453 * @size: wanted size in bytes 454 * @align: wanted align 455 * 456 * Try to allocate @size bytes area aligned at @align from @chunk. 457 * Note that this function only allocates the offset. It doesn't 458 * populate or map the area. 459 * 460 * @chunk->map must have at least two free slots. 461 * 462 * CONTEXT: 463 * pcpu_lock. 464 * 465 * RETURNS: 466 * Allocated offset in @chunk on success, -1 if no matching area is 467 * found. 468 */ 469static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) 470{ 471 int oslot = pcpu_chunk_slot(chunk); 472 int max_contig = 0; 473 int i, off; 474 475 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { 476 bool is_last = i + 1 == chunk->map_used; 477 int head, tail; 478 479 /* extra for alignment requirement */ 480 head = ALIGN(off, align) - off; 481 BUG_ON(i == 0 && head != 0); 482 483 if (chunk->map[i] < 0) 484 continue; 485 if (chunk->map[i] < head + size) { 486 max_contig = max(chunk->map[i], max_contig); 487 continue; 488 } 489 490 /* 491 * If head is small or the previous block is free, 492 * merge'em. Note that 'small' is defined as smaller 493 * than sizeof(int), which is very small but isn't too 494 * uncommon for percpu allocations. 495 */ 496 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { 497 if (chunk->map[i - 1] > 0) 498 chunk->map[i - 1] += head; 499 else { 500 chunk->map[i - 1] -= head; 501 chunk->free_size -= head; 502 } 503 chunk->map[i] -= head; 504 off += head; 505 head = 0; 506 } 507 508 /* if tail is small, just keep it around */ 509 tail = chunk->map[i] - head - size; 510 if (tail < sizeof(int)) 511 tail = 0; 512 513 /* split if warranted */ 514 if (head || tail) { 515 pcpu_split_block(chunk, i, head, tail); 516 if (head) { 517 i++; 518 off += head; 519 max_contig = max(chunk->map[i - 1], max_contig); 520 } 521 if (tail) 522 max_contig = max(chunk->map[i + 1], max_contig); 523 } 524 525 /* update hint and mark allocated */ 526 if (is_last) 527 chunk->contig_hint = max_contig; /* fully scanned */ 528 else 529 chunk->contig_hint = max(chunk->contig_hint, 530 max_contig); 531 532 chunk->free_size -= chunk->map[i]; 533 chunk->map[i] = -chunk->map[i]; 534 535 pcpu_chunk_relocate(chunk, oslot); 536 return off; 537 } 538 539 chunk->contig_hint = max_contig; /* fully scanned */ 540 pcpu_chunk_relocate(chunk, oslot); 541 542 /* tell the upper layer that this chunk has no matching area */ 543 return -1; 544} 545 546/** 547 * pcpu_free_area - free area to a pcpu_chunk 548 * @chunk: chunk of interest 549 * @freeme: offset of area to free 550 * 551 * Free area starting from @freeme to @chunk. Note that this function 552 * only modifies the allocation map. It doesn't depopulate or unmap 553 * the area. 554 * 555 * CONTEXT: 556 * pcpu_lock. 557 */ 558static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) 559{ 560 int oslot = pcpu_chunk_slot(chunk); 561 int i, off; 562 563 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) 564 if (off == freeme) 565 break; 566 BUG_ON(off != freeme); 567 BUG_ON(chunk->map[i] > 0); 568 569 chunk->map[i] = -chunk->map[i]; 570 chunk->free_size += chunk->map[i]; 571 572 /* merge with previous? */ 573 if (i > 0 && chunk->map[i - 1] >= 0) { 574 chunk->map[i - 1] += chunk->map[i]; 575 chunk->map_used--; 576 memmove(&chunk->map[i], &chunk->map[i + 1], 577 (chunk->map_used - i) * sizeof(chunk->map[0])); 578 i--; 579 } 580 /* merge with next? */ 581 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { 582 chunk->map[i] += chunk->map[i + 1]; 583 chunk->map_used--; 584 memmove(&chunk->map[i + 1], &chunk->map[i + 2], 585 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); 586 } 587 588 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); 589 pcpu_chunk_relocate(chunk, oslot); 590} 591 592/** 593 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap 594 * @chunk: chunk of interest 595 * @bitmapp: output parameter for bitmap 596 * @may_alloc: may allocate the array 597 * 598 * Returns pointer to array of pointers to struct page and bitmap, 599 * both of which can be indexed with pcpu_page_idx(). The returned 600 * array is cleared to zero and *@bitmapp is copied from 601 * @chunk->populated. Note that there is only one array and bitmap 602 * and access exclusion is the caller's responsibility. 603 * 604 * CONTEXT: 605 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc. 606 * Otherwise, don't care. 607 * 608 * RETURNS: 609 * Pointer to temp pages array on success, NULL on failure. 610 */ 611static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk, 612 unsigned long **bitmapp, 613 bool may_alloc) 614{ 615 static struct page **pages; 616 static unsigned long *bitmap; 617 size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]); 618 size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) * 619 sizeof(unsigned long); 620 621 if (!pages || !bitmap) { 622 if (may_alloc && !pages) 623 pages = pcpu_mem_alloc(pages_size); 624 if (may_alloc && !bitmap) 625 bitmap = pcpu_mem_alloc(bitmap_size); 626 if (!pages || !bitmap) 627 return NULL; 628 } 629 630 memset(pages, 0, pages_size); 631 bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages); 632 633 *bitmapp = bitmap; 634 return pages; 635} 636 637/** 638 * pcpu_free_pages - free pages which were allocated for @chunk 639 * @chunk: chunk pages were allocated for 640 * @pages: array of pages to be freed, indexed by pcpu_page_idx() 641 * @populated: populated bitmap 642 * @page_start: page index of the first page to be freed 643 * @page_end: page index of the last page to be freed + 1 644 * 645 * Free pages [@page_start and @page_end) in @pages for all units. 646 * The pages were allocated for @chunk. 647 */ 648static void pcpu_free_pages(struct pcpu_chunk *chunk, 649 struct page **pages, unsigned long *populated, 650 int page_start, int page_end) 651{ 652 unsigned int cpu; 653 int i; 654 655 for_each_possible_cpu(cpu) { 656 for (i = page_start; i < page_end; i++) { 657 struct page *page = pages[pcpu_page_idx(cpu, i)]; 658 659 if (page) 660 __free_page(page); 661 } 662 } 663} 664 665/** 666 * pcpu_alloc_pages - allocates pages for @chunk 667 * @chunk: target chunk 668 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx() 669 * @populated: populated bitmap 670 * @page_start: page index of the first page to be allocated 671 * @page_end: page index of the last page to be allocated + 1 672 * 673 * Allocate pages [@page_start,@page_end) into @pages for all units. 674 * The allocation is for @chunk. Percpu core doesn't care about the 675 * content of @pages and will pass it verbatim to pcpu_map_pages(). 676 */ 677static int pcpu_alloc_pages(struct pcpu_chunk *chunk, 678 struct page **pages, unsigned long *populated, 679 int page_start, int page_end) 680{ 681 const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; 682 unsigned int cpu; 683 int i; 684 685 for_each_possible_cpu(cpu) { 686 for (i = page_start; i < page_end; i++) { 687 struct page **pagep = &pages[pcpu_page_idx(cpu, i)]; 688 689 *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0); 690 if (!*pagep) { 691 pcpu_free_pages(chunk, pages, populated, 692 page_start, page_end); 693 return -ENOMEM; 694 } 695 } 696 } 697 return 0; 698} 699 700/** 701 * pcpu_pre_unmap_flush - flush cache prior to unmapping 702 * @chunk: chunk the regions to be flushed belongs to 703 * @page_start: page index of the first page to be flushed 704 * @page_end: page index of the last page to be flushed + 1 705 * 706 * Pages in [@page_start,@page_end) of @chunk are about to be 707 * unmapped. Flush cache. As each flushing trial can be very 708 * expensive, issue flush on the whole region at once rather than 709 * doing it for each cpu. This could be an overkill but is more 710 * scalable. 711 */ 712static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk, 713 int page_start, int page_end) 714{ 715 flush_cache_vunmap( 716 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), 717 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); 718} 719 720static void __pcpu_unmap_pages(unsigned long addr, int nr_pages) 721{ 722 unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT); 723} 724 725/** 726 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk 727 * @chunk: chunk of interest 728 * @pages: pages array which can be used to pass information to free 729 * @populated: populated bitmap 730 * @page_start: page index of the first page to unmap 731 * @page_end: page index of the last page to unmap + 1 732 * 733 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. 734 * Corresponding elements in @pages were cleared by the caller and can 735 * be used to carry information to pcpu_free_pages() which will be 736 * called after all unmaps are finished. The caller should call 737 * proper pre/post flush functions. 738 */ 739static void pcpu_unmap_pages(struct pcpu_chunk *chunk, 740 struct page **pages, unsigned long *populated, 741 int page_start, int page_end) 742{ 743 unsigned int cpu; 744 int i; 745 746 for_each_possible_cpu(cpu) { 747 for (i = page_start; i < page_end; i++) { 748 struct page *page; 749 750 page = pcpu_chunk_page(chunk, cpu, i); 751 WARN_ON(!page); 752 pages[pcpu_page_idx(cpu, i)] = page; 753 } 754 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start), 755 page_end - page_start); 756 } 757 758 for (i = page_start; i < page_end; i++) 759 __clear_bit(i, populated); 760} 761 762/** 763 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping 764 * @chunk: pcpu_chunk the regions to be flushed belong to 765 * @page_start: page index of the first page to be flushed 766 * @page_end: page index of the last page to be flushed + 1 767 * 768 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush 769 * TLB for the regions. This can be skipped if the area is to be 770 * returned to vmalloc as vmalloc will handle TLB flushing lazily. 771 * 772 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once 773 * for the whole region. 774 */ 775static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk, 776 int page_start, int page_end) 777{ 778 flush_tlb_kernel_range( 779 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), 780 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); 781} 782 783static int __pcpu_map_pages(unsigned long addr, struct page **pages, 784 int nr_pages) 785{ 786 return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT, 787 PAGE_KERNEL, pages); 788} 789 790/** 791 * pcpu_map_pages - map pages into a pcpu_chunk 792 * @chunk: chunk of interest 793 * @pages: pages array containing pages to be mapped 794 * @populated: populated bitmap 795 * @page_start: page index of the first page to map 796 * @page_end: page index of the last page to map + 1 797 * 798 * For each cpu, map pages [@page_start,@page_end) into @chunk. The 799 * caller is responsible for calling pcpu_post_map_flush() after all 800 * mappings are complete. 801 * 802 * This function is responsible for setting corresponding bits in 803 * @chunk->populated bitmap and whatever is necessary for reverse 804 * lookup (addr -> chunk). 805 */ 806static int pcpu_map_pages(struct pcpu_chunk *chunk, 807 struct page **pages, unsigned long *populated, 808 int page_start, int page_end) 809{ 810 unsigned int cpu, tcpu; 811 int i, err; 812 813 for_each_possible_cpu(cpu) { 814 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start), 815 &pages[pcpu_page_idx(cpu, page_start)], 816 page_end - page_start); 817 if (err < 0) 818 goto err; 819 } 820 821 /* mapping successful, link chunk and mark populated */ 822 for (i = page_start; i < page_end; i++) { 823 for_each_possible_cpu(cpu) 824 pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)], 825 chunk); 826 __set_bit(i, populated); 827 } 828 829 return 0; 830 831err: 832 for_each_possible_cpu(tcpu) { 833 if (tcpu == cpu) 834 break; 835 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start), 836 page_end - page_start); 837 } 838 return err; 839} 840 841/** 842 * pcpu_post_map_flush - flush cache after mapping 843 * @chunk: pcpu_chunk the regions to be flushed belong to 844 * @page_start: page index of the first page to be flushed 845 * @page_end: page index of the last page to be flushed + 1 846 * 847 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush 848 * cache. 849 * 850 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once 851 * for the whole region. 852 */ 853static void pcpu_post_map_flush(struct pcpu_chunk *chunk, 854 int page_start, int page_end) 855{ 856 flush_cache_vmap( 857 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), 858 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); 859} 860 861/** 862 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk 863 * @chunk: chunk to depopulate 864 * @off: offset to the area to depopulate 865 * @size: size of the area to depopulate in bytes 866 * @flush: whether to flush cache and tlb or not 867 * 868 * For each cpu, depopulate and unmap pages [@page_start,@page_end) 869 * from @chunk. If @flush is true, vcache is flushed before unmapping 870 * and tlb after. 871 * 872 * CONTEXT: 873 * pcpu_alloc_mutex. 874 */ 875static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size) 876{ 877 int page_start = PFN_DOWN(off); 878 int page_end = PFN_UP(off + size); 879 struct page **pages; 880 unsigned long *populated; 881 int rs, re; 882 883 /* quick path, check whether it's empty already */ 884 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { 885 if (rs == page_start && re == page_end) 886 return; 887 break; 888 } 889 890 /* immutable chunks can't be depopulated */ 891 WARN_ON(chunk->immutable); 892 893 /* 894 * If control reaches here, there must have been at least one 895 * successful population attempt so the temp pages array must 896 * be available now. 897 */ 898 pages = pcpu_get_pages_and_bitmap(chunk, &populated, false); 899 BUG_ON(!pages); 900 901 /* unmap and free */ 902 pcpu_pre_unmap_flush(chunk, page_start, page_end); 903 904 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) 905 pcpu_unmap_pages(chunk, pages, populated, rs, re); 906 907 /* no need to flush tlb, vmalloc will handle it lazily */ 908 909 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) 910 pcpu_free_pages(chunk, pages, populated, rs, re); 911 912 /* commit new bitmap */ 913 bitmap_copy(chunk->populated, populated, pcpu_unit_pages); 914} 915 916/** 917 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk 918 * @chunk: chunk of interest 919 * @off: offset to the area to populate 920 * @size: size of the area to populate in bytes 921 * 922 * For each cpu, populate and map pages [@page_start,@page_end) into 923 * @chunk. The area is cleared on return. 924 * 925 * CONTEXT: 926 * pcpu_alloc_mutex, does GFP_KERNEL allocation. 927 */ 928static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) 929{ 930 int page_start = PFN_DOWN(off); 931 int page_end = PFN_UP(off + size); 932 int free_end = page_start, unmap_end = page_start; 933 struct page **pages; 934 unsigned long *populated; 935 unsigned int cpu; 936 int rs, re, rc; 937 938 /* quick path, check whether all pages are already there */ 939 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) { 940 if (rs == page_start && re == page_end) 941 goto clear; 942 break; 943 } 944 945 /* need to allocate and map pages, this chunk can't be immutable */ 946 WARN_ON(chunk->immutable); 947 948 pages = pcpu_get_pages_and_bitmap(chunk, &populated, true); 949 if (!pages) 950 return -ENOMEM; 951 952 /* alloc and map */ 953 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { 954 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re); 955 if (rc) 956 goto err_free; 957 free_end = re; 958 } 959 960 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { 961 rc = pcpu_map_pages(chunk, pages, populated, rs, re); 962 if (rc) 963 goto err_unmap; 964 unmap_end = re; 965 } 966 pcpu_post_map_flush(chunk, page_start, page_end); 967 968 /* commit new bitmap */ 969 bitmap_copy(chunk->populated, populated, pcpu_unit_pages); 970clear: 971 for_each_possible_cpu(cpu) 972 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); 973 return 0; 974 975err_unmap: 976 pcpu_pre_unmap_flush(chunk, page_start, unmap_end); 977 pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end) 978 pcpu_unmap_pages(chunk, pages, populated, rs, re); 979 pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end); 980err_free: 981 pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end) 982 pcpu_free_pages(chunk, pages, populated, rs, re); 983 return rc; 984} 985 986static void free_pcpu_chunk(struct pcpu_chunk *chunk) 987{ 988 if (!chunk) 989 return; 990 if (chunk->vm) 991 free_vm_area(chunk->vm); 992 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); 993 kfree(chunk); 994} 995 996static struct pcpu_chunk *alloc_pcpu_chunk(void) 997{ 998 struct pcpu_chunk *chunk; 999 1000 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); 1001 if (!chunk) 1002 return NULL; 1003 1004 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); 1005 chunk->map_alloc = PCPU_DFL_MAP_ALLOC; 1006 chunk->map[chunk->map_used++] = pcpu_unit_size; 1007 1008 chunk->vm = get_vm_area(pcpu_chunk_size, VM_ALLOC); 1009 if (!chunk->vm) { 1010 free_pcpu_chunk(chunk); 1011 return NULL; 1012 } 1013 1014 INIT_LIST_HEAD(&chunk->list); 1015 chunk->free_size = pcpu_unit_size; 1016 chunk->contig_hint = pcpu_unit_size; 1017 1018 return chunk; 1019} 1020 1021/** 1022 * pcpu_alloc - the percpu allocator 1023 * @size: size of area to allocate in bytes 1024 * @align: alignment of area (max PAGE_SIZE) 1025 * @reserved: allocate from the reserved chunk if available 1026 * 1027 * Allocate percpu area of @size bytes aligned at @align. 1028 * 1029 * CONTEXT: 1030 * Does GFP_KERNEL allocation. 1031 * 1032 * RETURNS: 1033 * Percpu pointer to the allocated area on success, NULL on failure. 1034 */ 1035static void *pcpu_alloc(size_t size, size_t align, bool reserved) 1036{ 1037 struct pcpu_chunk *chunk; 1038 int slot, off; 1039 1040 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { 1041 WARN(true, "illegal size (%zu) or align (%zu) for " 1042 "percpu allocation\n", size, align); 1043 return NULL; 1044 } 1045 1046 mutex_lock(&pcpu_alloc_mutex); 1047 spin_lock_irq(&pcpu_lock); 1048 1049 /* serve reserved allocations from the reserved chunk if available */ 1050 if (reserved && pcpu_reserved_chunk) { 1051 chunk = pcpu_reserved_chunk; 1052 if (size > chunk->contig_hint || 1053 pcpu_extend_area_map(chunk) < 0) 1054 goto fail_unlock; 1055 off = pcpu_alloc_area(chunk, size, align); 1056 if (off >= 0) 1057 goto area_found; 1058 goto fail_unlock; 1059 } 1060 1061restart: 1062 /* search through normal chunks */ 1063 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 1064 list_for_each_entry(chunk, &pcpu_slot[slot], list) { 1065 if (size > chunk->contig_hint) 1066 continue; 1067 1068 switch (pcpu_extend_area_map(chunk)) { 1069 case 0: 1070 break; 1071 case 1: 1072 goto restart; /* pcpu_lock dropped, restart */ 1073 default: 1074 goto fail_unlock; 1075 } 1076 1077 off = pcpu_alloc_area(chunk, size, align); 1078 if (off >= 0) 1079 goto area_found; 1080 } 1081 } 1082 1083 /* hmmm... no space left, create a new chunk */ 1084 spin_unlock_irq(&pcpu_lock); 1085 1086 chunk = alloc_pcpu_chunk(); 1087 if (!chunk) 1088 goto fail_unlock_mutex; 1089 1090 spin_lock_irq(&pcpu_lock); 1091 pcpu_chunk_relocate(chunk, -1); 1092 goto restart; 1093 1094area_found: 1095 spin_unlock_irq(&pcpu_lock); 1096 1097 /* populate, map and clear the area */ 1098 if (pcpu_populate_chunk(chunk, off, size)) { 1099 spin_lock_irq(&pcpu_lock); 1100 pcpu_free_area(chunk, off); 1101 goto fail_unlock; 1102 } 1103 1104 mutex_unlock(&pcpu_alloc_mutex); 1105 1106 /* return address relative to unit0 */ 1107 return __addr_to_pcpu_ptr(chunk->vm->addr + off); 1108 1109fail_unlock: 1110 spin_unlock_irq(&pcpu_lock); 1111fail_unlock_mutex: 1112 mutex_unlock(&pcpu_alloc_mutex); 1113 return NULL; 1114} 1115 1116/** 1117 * __alloc_percpu - allocate dynamic percpu area 1118 * @size: size of area to allocate in bytes 1119 * @align: alignment of area (max PAGE_SIZE) 1120 * 1121 * Allocate percpu area of @size bytes aligned at @align. Might 1122 * sleep. Might trigger writeouts. 1123 * 1124 * CONTEXT: 1125 * Does GFP_KERNEL allocation. 1126 * 1127 * RETURNS: 1128 * Percpu pointer to the allocated area on success, NULL on failure. 1129 */ 1130void *__alloc_percpu(size_t size, size_t align) 1131{ 1132 return pcpu_alloc(size, align, false); 1133} 1134EXPORT_SYMBOL_GPL(__alloc_percpu); 1135 1136/** 1137 * __alloc_reserved_percpu - allocate reserved percpu area 1138 * @size: size of area to allocate in bytes 1139 * @align: alignment of area (max PAGE_SIZE) 1140 * 1141 * Allocate percpu area of @size bytes aligned at @align from reserved 1142 * percpu area if arch has set it up; otherwise, allocation is served 1143 * from the same dynamic area. Might sleep. Might trigger writeouts. 1144 * 1145 * CONTEXT: 1146 * Does GFP_KERNEL allocation. 1147 * 1148 * RETURNS: 1149 * Percpu pointer to the allocated area on success, NULL on failure. 1150 */ 1151void *__alloc_reserved_percpu(size_t size, size_t align) 1152{ 1153 return pcpu_alloc(size, align, true); 1154} 1155 1156/** 1157 * pcpu_reclaim - reclaim fully free chunks, workqueue function 1158 * @work: unused 1159 * 1160 * Reclaim all fully free chunks except for the first one. 1161 * 1162 * CONTEXT: 1163 * workqueue context. 1164 */ 1165static void pcpu_reclaim(struct work_struct *work) 1166{ 1167 LIST_HEAD(todo); 1168 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; 1169 struct pcpu_chunk *chunk, *next; 1170 1171 mutex_lock(&pcpu_alloc_mutex); 1172 spin_lock_irq(&pcpu_lock); 1173 1174 list_for_each_entry_safe(chunk, next, head, list) { 1175 WARN_ON(chunk->immutable); 1176 1177 /* spare the first one */ 1178 if (chunk == list_first_entry(head, struct pcpu_chunk, list)) 1179 continue; 1180 1181 list_move(&chunk->list, &todo); 1182 } 1183 1184 spin_unlock_irq(&pcpu_lock); 1185 1186 list_for_each_entry_safe(chunk, next, &todo, list) { 1187 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size); 1188 free_pcpu_chunk(chunk); 1189 } 1190 1191 mutex_unlock(&pcpu_alloc_mutex); 1192} 1193 1194/** 1195 * free_percpu - free percpu area 1196 * @ptr: pointer to area to free 1197 * 1198 * Free percpu area @ptr. 1199 * 1200 * CONTEXT: 1201 * Can be called from atomic context. 1202 */ 1203void free_percpu(void *ptr) 1204{ 1205 void *addr = __pcpu_ptr_to_addr(ptr); 1206 struct pcpu_chunk *chunk; 1207 unsigned long flags; 1208 int off; 1209 1210 if (!ptr) 1211 return; 1212 1213 spin_lock_irqsave(&pcpu_lock, flags); 1214 1215 chunk = pcpu_chunk_addr_search(addr); 1216 off = addr - chunk->vm->addr; 1217 1218 pcpu_free_area(chunk, off); 1219 1220 /* if there are more than one fully free chunks, wake up grim reaper */ 1221 if (chunk->free_size == pcpu_unit_size) { 1222 struct pcpu_chunk *pos; 1223 1224 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 1225 if (pos != chunk) { 1226 schedule_work(&pcpu_reclaim_work); 1227 break; 1228 } 1229 } 1230 1231 spin_unlock_irqrestore(&pcpu_lock, flags); 1232} 1233EXPORT_SYMBOL_GPL(free_percpu); 1234 1235static inline size_t pcpu_calc_fc_sizes(size_t static_size, 1236 size_t reserved_size, 1237 ssize_t *dyn_sizep) 1238{ 1239 size_t size_sum; 1240 1241 size_sum = PFN_ALIGN(static_size + reserved_size + 1242 (*dyn_sizep >= 0 ? *dyn_sizep : 0)); 1243 if (*dyn_sizep != 0) 1244 *dyn_sizep = size_sum - static_size - reserved_size; 1245 1246 return size_sum; 1247} 1248 1249/** 1250 * pcpu_alloc_alloc_info - allocate percpu allocation info 1251 * @nr_groups: the number of groups 1252 * @nr_units: the number of units 1253 * 1254 * Allocate ai which is large enough for @nr_groups groups containing 1255 * @nr_units units. The returned ai's groups[0].cpu_map points to the 1256 * cpu_map array which is long enough for @nr_units and filled with 1257 * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1258 * pointer of other groups. 1259 * 1260 * RETURNS: 1261 * Pointer to the allocated pcpu_alloc_info on success, NULL on 1262 * failure. 1263 */ 1264struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 1265 int nr_units) 1266{ 1267 struct pcpu_alloc_info *ai; 1268 size_t base_size, ai_size; 1269 void *ptr; 1270 int unit; 1271 1272 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 1273 __alignof__(ai->groups[0].cpu_map[0])); 1274 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 1275 1276 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size)); 1277 if (!ptr) 1278 return NULL; 1279 ai = ptr; 1280 ptr += base_size; 1281 1282 ai->groups[0].cpu_map = ptr; 1283 1284 for (unit = 0; unit < nr_units; unit++) 1285 ai->groups[0].cpu_map[unit] = NR_CPUS; 1286 1287 ai->nr_groups = nr_groups; 1288 ai->__ai_size = PFN_ALIGN(ai_size); 1289 1290 return ai; 1291} 1292 1293/** 1294 * pcpu_free_alloc_info - free percpu allocation info 1295 * @ai: pcpu_alloc_info to free 1296 * 1297 * Free @ai which was allocated by pcpu_alloc_alloc_info(). 1298 */ 1299void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 1300{ 1301 free_bootmem(__pa(ai), ai->__ai_size); 1302} 1303 1304/** 1305 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 1306 * @reserved_size: the size of reserved percpu area in bytes 1307 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto 1308 * @atom_size: allocation atom size 1309 * @cpu_distance_fn: callback to determine distance between cpus, optional 1310 * 1311 * This function determines grouping of units, their mappings to cpus 1312 * and other parameters considering needed percpu size, allocation 1313 * atom size and distances between CPUs. 1314 * 1315 * Groups are always mutliples of atom size and CPUs which are of 1316 * LOCAL_DISTANCE both ways are grouped together and share space for 1317 * units in the same group. The returned configuration is guaranteed 1318 * to have CPUs on different nodes on different groups and >=75% usage 1319 * of allocated virtual address space. 1320 * 1321 * RETURNS: 1322 * On success, pointer to the new allocation_info is returned. On 1323 * failure, ERR_PTR value is returned. 1324 */ 1325struct pcpu_alloc_info * __init pcpu_build_alloc_info( 1326 size_t reserved_size, ssize_t dyn_size, 1327 size_t atom_size, 1328 pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 1329{ 1330 static int group_map[NR_CPUS] __initdata; 1331 static int group_cnt[NR_CPUS] __initdata; 1332 const size_t static_size = __per_cpu_end - __per_cpu_start; 1333 int group_cnt_max = 0, nr_groups = 1, nr_units = 0; 1334 size_t size_sum, min_unit_size, alloc_size; 1335 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 1336 int last_allocs, group, unit; 1337 unsigned int cpu, tcpu; 1338 struct pcpu_alloc_info *ai; 1339 unsigned int *cpu_map; 1340 1341 /* 1342 * Determine min_unit_size, alloc_size and max_upa such that 1343 * alloc_size is multiple of atom_size and is the smallest 1344 * which can accomodate 4k aligned segments which are equal to 1345 * or larger than min_unit_size. 1346 */ 1347 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size); 1348 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 1349 1350 alloc_size = roundup(min_unit_size, atom_size); 1351 upa = alloc_size / min_unit_size; 1352 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1353 upa--; 1354 max_upa = upa; 1355 1356 /* group cpus according to their proximity */ 1357 for_each_possible_cpu(cpu) { 1358 group = 0; 1359 next_group: 1360 for_each_possible_cpu(tcpu) { 1361 if (cpu == tcpu) 1362 break; 1363 if (group_map[tcpu] == group && cpu_distance_fn && 1364 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 1365 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 1366 group++; 1367 nr_groups = max(nr_groups, group + 1); 1368 goto next_group; 1369 } 1370 } 1371 group_map[cpu] = group; 1372 group_cnt[group]++; 1373 group_cnt_max = max(group_cnt_max, group_cnt[group]); 1374 } 1375 1376 /* 1377 * Expand unit size until address space usage goes over 75% 1378 * and then as much as possible without using more address 1379 * space. 1380 */ 1381 last_allocs = INT_MAX; 1382 for (upa = max_upa; upa; upa--) { 1383 int allocs = 0, wasted = 0; 1384 1385 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1386 continue; 1387 1388 for (group = 0; group < nr_groups; group++) { 1389 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 1390 allocs += this_allocs; 1391 wasted += this_allocs * upa - group_cnt[group]; 1392 } 1393 1394 /* 1395 * Don't accept if wastage is over 25%. The 1396 * greater-than comparison ensures upa==1 always 1397 * passes the following check. 1398 */ 1399 if (wasted > num_possible_cpus() / 3) 1400 continue; 1401 1402 /* and then don't consume more memory */ 1403 if (allocs > last_allocs) 1404 break; 1405 last_allocs = allocs; 1406 best_upa = upa; 1407 } 1408 upa = best_upa; 1409 1410 /* allocate and fill alloc_info */ 1411 for (group = 0; group < nr_groups; group++) 1412 nr_units += roundup(group_cnt[group], upa); 1413 1414 ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 1415 if (!ai) 1416 return ERR_PTR(-ENOMEM); 1417 cpu_map = ai->groups[0].cpu_map; 1418 1419 for (group = 0; group < nr_groups; group++) { 1420 ai->groups[group].cpu_map = cpu_map; 1421 cpu_map += roundup(group_cnt[group], upa); 1422 } 1423 1424 ai->static_size = static_size; 1425 ai->reserved_size = reserved_size; 1426 ai->dyn_size = dyn_size; 1427 ai->unit_size = alloc_size / upa; 1428 ai->atom_size = atom_size; 1429 ai->alloc_size = alloc_size; 1430 1431 for (group = 0, unit = 0; group_cnt[group]; group++) { 1432 struct pcpu_group_info *gi = &ai->groups[group]; 1433 1434 /* 1435 * Initialize base_offset as if all groups are located 1436 * back-to-back. The caller should update this to 1437 * reflect actual allocation. 1438 */ 1439 gi->base_offset = unit * ai->unit_size; 1440 1441 for_each_possible_cpu(cpu) 1442 if (group_map[cpu] == group) 1443 gi->cpu_map[gi->nr_units++] = cpu; 1444 gi->nr_units = roundup(gi->nr_units, upa); 1445 unit += gi->nr_units; 1446 } 1447 BUG_ON(unit != nr_units); 1448 1449 return ai; 1450} 1451 1452/** 1453 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 1454 * @lvl: loglevel 1455 * @ai: allocation info to dump 1456 * 1457 * Print out information about @ai using loglevel @lvl. 1458 */ 1459static void pcpu_dump_alloc_info(const char *lvl, 1460 const struct pcpu_alloc_info *ai) 1461{ 1462 int group_width = 1, cpu_width = 1, width; 1463 char empty_str[] = "--------"; 1464 int alloc = 0, alloc_end = 0; 1465 int group, v; 1466 int upa, apl; /* units per alloc, allocs per line */ 1467 1468 v = ai->nr_groups; 1469 while (v /= 10) 1470 group_width++; 1471 1472 v = num_possible_cpus(); 1473 while (v /= 10) 1474 cpu_width++; 1475 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 1476 1477 upa = ai->alloc_size / ai->unit_size; 1478 width = upa * (cpu_width + 1) + group_width + 3; 1479 apl = rounddown_pow_of_two(max(60 / width, 1)); 1480 1481 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 1482 lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 1483 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 1484 1485 for (group = 0; group < ai->nr_groups; group++) { 1486 const struct pcpu_group_info *gi = &ai->groups[group]; 1487 int unit = 0, unit_end = 0; 1488 1489 BUG_ON(gi->nr_units % upa); 1490 for (alloc_end += gi->nr_units / upa; 1491 alloc < alloc_end; alloc++) { 1492 if (!(alloc % apl)) { 1493 printk("\n"); 1494 printk("%spcpu-alloc: ", lvl); 1495 } 1496 printk("[%0*d] ", group_width, group); 1497 1498 for (unit_end += upa; unit < unit_end; unit++) 1499 if (gi->cpu_map[unit] != NR_CPUS) 1500 printk("%0*d ", cpu_width, 1501 gi->cpu_map[unit]); 1502 else 1503 printk("%s ", empty_str); 1504 } 1505 } 1506 printk("\n"); 1507} 1508 1509/** 1510 * pcpu_setup_first_chunk - initialize the first percpu chunk 1511 * @ai: pcpu_alloc_info describing how to percpu area is shaped 1512 * @base_addr: mapped address 1513 * 1514 * Initialize the first percpu chunk which contains the kernel static 1515 * perpcu area. This function is to be called from arch percpu area 1516 * setup path. 1517 * 1518 * @ai contains all information necessary to initialize the first 1519 * chunk and prime the dynamic percpu allocator. 1520 * 1521 * @ai->static_size is the size of static percpu area. 1522 * 1523 * @ai->reserved_size, if non-zero, specifies the amount of bytes to 1524 * reserve after the static area in the first chunk. This reserves 1525 * the first chunk such that it's available only through reserved 1526 * percpu allocation. This is primarily used to serve module percpu 1527 * static areas on architectures where the addressing model has 1528 * limited offset range for symbol relocations to guarantee module 1529 * percpu symbols fall inside the relocatable range. 1530 * 1531 * @ai->dyn_size determines the number of bytes available for dynamic 1532 * allocation in the first chunk. The area between @ai->static_size + 1533 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 1534 * 1535 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 1536 * and equal to or larger than @ai->static_size + @ai->reserved_size + 1537 * @ai->dyn_size. 1538 * 1539 * @ai->atom_size is the allocation atom size and used as alignment 1540 * for vm areas. 1541 * 1542 * @ai->alloc_size is the allocation size and always multiple of 1543 * @ai->atom_size. This is larger than @ai->atom_size if 1544 * @ai->unit_size is larger than @ai->atom_size. 1545 * 1546 * @ai->nr_groups and @ai->groups describe virtual memory layout of 1547 * percpu areas. Units which should be colocated are put into the 1548 * same group. Dynamic VM areas will be allocated according to these 1549 * groupings. If @ai->nr_groups is zero, a single group containing 1550 * all units is assumed. 1551 * 1552 * The caller should have mapped the first chunk at @base_addr and 1553 * copied static data to each unit. 1554 * 1555 * If the first chunk ends up with both reserved and dynamic areas, it 1556 * is served by two chunks - one to serve the core static and reserved 1557 * areas and the other for the dynamic area. They share the same vm 1558 * and page map but uses different area allocation map to stay away 1559 * from each other. The latter chunk is circulated in the chunk slots 1560 * and available for dynamic allocation like any other chunks. 1561 * 1562 * RETURNS: 1563 * 0 on success, -errno on failure. 1564 */ 1565int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 1566 void *base_addr) 1567{ 1568 static struct vm_struct first_vm; 1569 static int smap[2], dmap[2]; 1570 size_t dyn_size = ai->dyn_size; 1571 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; 1572 struct pcpu_chunk *schunk, *dchunk = NULL; 1573 unsigned long *unit_off; 1574 unsigned int cpu; 1575 int *unit_map; 1576 int group, unit, i; 1577 1578 /* sanity checks */ 1579 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || 1580 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); 1581 BUG_ON(ai->nr_groups <= 0); 1582 BUG_ON(!ai->static_size); 1583 BUG_ON(!base_addr); 1584 BUG_ON(ai->unit_size < size_sum); 1585 BUG_ON(ai->unit_size & ~PAGE_MASK); 1586 BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 1587 1588 pcpu_dump_alloc_info(KERN_DEBUG, ai); 1589 1590 /* determine number of units and initialize unit_map and base */ 1591 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0])); 1592 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0])); 1593 1594 for (cpu = 0; cpu < nr_cpu_ids; cpu++) 1595 unit_map[cpu] = NR_CPUS; 1596 pcpu_first_unit_cpu = NR_CPUS; 1597 1598 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 1599 const struct pcpu_group_info *gi = &ai->groups[group]; 1600 1601 for (i = 0; i < gi->nr_units; i++) { 1602 cpu = gi->cpu_map[i]; 1603 if (cpu == NR_CPUS) 1604 continue; 1605 1606 BUG_ON(cpu > nr_cpu_ids || !cpu_possible(cpu)); 1607 BUG_ON(unit_map[cpu] != NR_CPUS); 1608 1609 unit_map[cpu] = unit + i; 1610 unit_off[cpu] = gi->base_offset + i * ai->unit_size; 1611 1612 if (pcpu_first_unit_cpu == NR_CPUS) 1613 pcpu_first_unit_cpu = cpu; 1614 } 1615 } 1616 pcpu_last_unit_cpu = cpu; 1617 pcpu_nr_units = unit; 1618 1619 for_each_possible_cpu(cpu) 1620 BUG_ON(unit_map[cpu] == NR_CPUS); 1621 1622 pcpu_unit_map = unit_map; 1623 pcpu_unit_offsets = unit_off; 1624 1625 /* determine basic parameters */ 1626 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 1627 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 1628 pcpu_chunk_size = pcpu_nr_units * pcpu_unit_size; 1629 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 1630 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 1631 1632 first_vm.flags = VM_ALLOC; 1633 first_vm.size = pcpu_chunk_size; 1634 first_vm.addr = base_addr; 1635 1636 /* 1637 * Allocate chunk slots. The additional last slot is for 1638 * empty chunks. 1639 */ 1640 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 1641 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); 1642 for (i = 0; i < pcpu_nr_slots; i++) 1643 INIT_LIST_HEAD(&pcpu_slot[i]); 1644 1645 /* 1646 * Initialize static chunk. If reserved_size is zero, the 1647 * static chunk covers static area + dynamic allocation area 1648 * in the first chunk. If reserved_size is not zero, it 1649 * covers static area + reserved area (mostly used for module 1650 * static percpu allocation). 1651 */ 1652 schunk = alloc_bootmem(pcpu_chunk_struct_size); 1653 INIT_LIST_HEAD(&schunk->list); 1654 schunk->vm = &first_vm; 1655 schunk->map = smap; 1656 schunk->map_alloc = ARRAY_SIZE(smap); 1657 schunk->immutable = true; 1658 bitmap_fill(schunk->populated, pcpu_unit_pages); 1659 1660 if (ai->reserved_size) { 1661 schunk->free_size = ai->reserved_size; 1662 pcpu_reserved_chunk = schunk; 1663 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; 1664 } else { 1665 schunk->free_size = dyn_size; 1666 dyn_size = 0; /* dynamic area covered */ 1667 } 1668 schunk->contig_hint = schunk->free_size; 1669 1670 schunk->map[schunk->map_used++] = -ai->static_size; 1671 if (schunk->free_size) 1672 schunk->map[schunk->map_used++] = schunk->free_size; 1673 1674 /* init dynamic chunk if necessary */ 1675 if (dyn_size) { 1676 dchunk = alloc_bootmem(pcpu_chunk_struct_size); 1677 INIT_LIST_HEAD(&dchunk->list); 1678 dchunk->vm = &first_vm; 1679 dchunk->map = dmap; 1680 dchunk->map_alloc = ARRAY_SIZE(dmap); 1681 dchunk->immutable = true; 1682 bitmap_fill(dchunk->populated, pcpu_unit_pages); 1683 1684 dchunk->contig_hint = dchunk->free_size = dyn_size; 1685 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; 1686 dchunk->map[dchunk->map_used++] = dchunk->free_size; 1687 } 1688 1689 /* link the first chunk in */ 1690 pcpu_first_chunk = dchunk ?: schunk; 1691 pcpu_chunk_relocate(pcpu_first_chunk, -1); 1692 1693 /* we're done */ 1694 pcpu_base_addr = schunk->vm->addr; 1695 return 0; 1696} 1697 1698const char *pcpu_fc_names[PCPU_FC_NR] __initdata = { 1699 [PCPU_FC_AUTO] = "auto", 1700 [PCPU_FC_EMBED] = "embed", 1701 [PCPU_FC_PAGE] = "page", 1702 [PCPU_FC_LPAGE] = "lpage", 1703}; 1704 1705enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 1706 1707static int __init percpu_alloc_setup(char *str) 1708{ 1709 if (0) 1710 /* nada */; 1711#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 1712 else if (!strcmp(str, "embed")) 1713 pcpu_chosen_fc = PCPU_FC_EMBED; 1714#endif 1715#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1716 else if (!strcmp(str, "page")) 1717 pcpu_chosen_fc = PCPU_FC_PAGE; 1718#endif 1719#ifdef CONFIG_NEED_PER_CPU_LPAGE_FIRST_CHUNK 1720 else if (!strcmp(str, "lpage")) 1721 pcpu_chosen_fc = PCPU_FC_LPAGE; 1722#endif 1723 else 1724 pr_warning("PERCPU: unknown allocator %s specified\n", str); 1725 1726 return 0; 1727} 1728early_param("percpu_alloc", percpu_alloc_setup); 1729 1730#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 1731 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 1732/** 1733 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 1734 * @reserved_size: the size of reserved percpu area in bytes 1735 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto 1736 * 1737 * This is a helper to ease setting up embedded first percpu chunk and 1738 * can be called where pcpu_setup_first_chunk() is expected. 1739 * 1740 * If this function is used to setup the first chunk, it is allocated 1741 * as a contiguous area using bootmem allocator and used as-is without 1742 * being mapped into vmalloc area. This enables the first chunk to 1743 * piggy back on the linear physical mapping which often uses larger 1744 * page size. 1745 * 1746 * When @dyn_size is positive, dynamic area might be larger than 1747 * specified to fill page alignment. When @dyn_size is auto, 1748 * @dyn_size is just big enough to fill page alignment after static 1749 * and reserved areas. 1750 * 1751 * If the needed size is smaller than the minimum or specified unit 1752 * size, the leftover is returned to the bootmem allocator. 1753 * 1754 * RETURNS: 1755 * 0 on success, -errno on failure. 1756 */ 1757int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size) 1758{ 1759 struct pcpu_alloc_info *ai; 1760 size_t size_sum, chunk_size; 1761 void *base; 1762 int unit; 1763 int rc; 1764 1765 ai = pcpu_build_alloc_info(reserved_size, dyn_size, PAGE_SIZE, NULL); 1766 if (IS_ERR(ai)) 1767 return PTR_ERR(ai); 1768 BUG_ON(ai->nr_groups != 1); 1769 BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); 1770 1771 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 1772 chunk_size = ai->unit_size * num_possible_cpus(); 1773 1774 base = __alloc_bootmem_nopanic(chunk_size, PAGE_SIZE, 1775 __pa(MAX_DMA_ADDRESS)); 1776 if (!base) { 1777 pr_warning("PERCPU: failed to allocate %zu bytes for " 1778 "embedding\n", chunk_size); 1779 rc = -ENOMEM; 1780 goto out_free_ai; 1781 } 1782 1783 /* return the leftover and copy */ 1784 for (unit = 0; unit < num_possible_cpus(); unit++) { 1785 void *ptr = base + unit * ai->unit_size; 1786 1787 free_bootmem(__pa(ptr + size_sum), ai->unit_size - size_sum); 1788 memcpy(ptr, __per_cpu_load, ai->static_size); 1789 } 1790 1791 /* we're ready, commit */ 1792 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 1793 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 1794 ai->dyn_size, ai->unit_size); 1795 1796 rc = pcpu_setup_first_chunk(ai, base); 1797out_free_ai: 1798 pcpu_free_alloc_info(ai); 1799 return rc; 1800} 1801#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK || 1802 !CONFIG_HAVE_SETUP_PER_CPU_AREA */ 1803 1804#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1805/** 1806 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 1807 * @reserved_size: the size of reserved percpu area in bytes 1808 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 1809 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE 1810 * @populate_pte_fn: function to populate pte 1811 * 1812 * This is a helper to ease setting up page-remapped first percpu 1813 * chunk and can be called where pcpu_setup_first_chunk() is expected. 1814 * 1815 * This is the basic allocator. Static percpu area is allocated 1816 * page-by-page into vmalloc area. 1817 * 1818 * RETURNS: 1819 * 0 on success, -errno on failure. 1820 */ 1821int __init pcpu_page_first_chunk(size_t reserved_size, 1822 pcpu_fc_alloc_fn_t alloc_fn, 1823 pcpu_fc_free_fn_t free_fn, 1824 pcpu_fc_populate_pte_fn_t populate_pte_fn) 1825{ 1826 static struct vm_struct vm; 1827 struct pcpu_alloc_info *ai; 1828 char psize_str[16]; 1829 int unit_pages; 1830 size_t pages_size; 1831 struct page **pages; 1832 int unit, i, j, rc; 1833 1834 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 1835 1836 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL); 1837 if (IS_ERR(ai)) 1838 return PTR_ERR(ai); 1839 BUG_ON(ai->nr_groups != 1); 1840 BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); 1841 1842 unit_pages = ai->unit_size >> PAGE_SHIFT; 1843 1844 /* unaligned allocations can't be freed, round up to page size */ 1845 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 1846 sizeof(pages[0])); 1847 pages = alloc_bootmem(pages_size); 1848 1849 /* allocate pages */ 1850 j = 0; 1851 for (unit = 0; unit < num_possible_cpus(); unit++) 1852 for (i = 0; i < unit_pages; i++) { 1853 unsigned int cpu = ai->groups[0].cpu_map[unit]; 1854 void *ptr; 1855 1856 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 1857 if (!ptr) { 1858 pr_warning("PERCPU: failed to allocate %s page " 1859 "for cpu%u\n", psize_str, cpu); 1860 goto enomem; 1861 } 1862 pages[j++] = virt_to_page(ptr); 1863 } 1864 1865 /* allocate vm area, map the pages and copy static data */ 1866 vm.flags = VM_ALLOC; 1867 vm.size = num_possible_cpus() * ai->unit_size; 1868 vm_area_register_early(&vm, PAGE_SIZE); 1869 1870 for (unit = 0; unit < num_possible_cpus(); unit++) { 1871 unsigned long unit_addr = 1872 (unsigned long)vm.addr + unit * ai->unit_size; 1873 1874 for (i = 0; i < unit_pages; i++) 1875 populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 1876 1877 /* pte already populated, the following shouldn't fail */ 1878 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 1879 unit_pages); 1880 if (rc < 0) 1881 panic("failed to map percpu area, err=%d\n", rc); 1882 1883 /* 1884 * FIXME: Archs with virtual cache should flush local 1885 * cache for the linear mapping here - something 1886 * equivalent to flush_cache_vmap() on the local cpu. 1887 * flush_cache_vmap() can't be used as most supporting 1888 * data structures are not set up yet. 1889 */ 1890 1891 /* copy static data */ 1892 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 1893 } 1894 1895 /* we're ready, commit */ 1896 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n", 1897 unit_pages, psize_str, vm.addr, ai->static_size, 1898 ai->reserved_size, ai->dyn_size); 1899 1900 rc = pcpu_setup_first_chunk(ai, vm.addr); 1901 goto out_free_ar; 1902 1903enomem: 1904 while (--j >= 0) 1905 free_fn(page_address(pages[j]), PAGE_SIZE); 1906 rc = -ENOMEM; 1907out_free_ar: 1908 free_bootmem(__pa(pages), pages_size); 1909 pcpu_free_alloc_info(ai); 1910 return rc; 1911} 1912#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */ 1913 1914#ifdef CONFIG_NEED_PER_CPU_LPAGE_FIRST_CHUNK 1915struct pcpul_ent { 1916 void *ptr; 1917 void *map_addr; 1918}; 1919 1920static size_t pcpul_size; 1921static size_t pcpul_lpage_size; 1922static int pcpul_nr_lpages; 1923static struct pcpul_ent *pcpul_map; 1924 1925static bool __init pcpul_unit_to_cpu(int unit, const struct pcpu_alloc_info *ai, 1926 unsigned int *cpup) 1927{ 1928 int group, cunit; 1929 1930 for (group = 0, cunit = 0; group < ai->nr_groups; group++) { 1931 const struct pcpu_group_info *gi = &ai->groups[group]; 1932 1933 if (unit < cunit + gi->nr_units) { 1934 if (cpup) 1935 *cpup = gi->cpu_map[unit - cunit]; 1936 return true; 1937 } 1938 cunit += gi->nr_units; 1939 } 1940 1941 return false; 1942} 1943 1944static int __init pcpul_cpu_to_unit(int cpu, const struct pcpu_alloc_info *ai) 1945{ 1946 int group, unit, i; 1947 1948 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 1949 const struct pcpu_group_info *gi = &ai->groups[group]; 1950 1951 for (i = 0; i < gi->nr_units; i++) 1952 if (gi->cpu_map[i] == cpu) 1953 return unit + i; 1954 } 1955 BUG(); 1956} 1957 1958/** 1959 * pcpu_lpage_first_chunk - remap the first percpu chunk using large page 1960 * @ai: pcpu_alloc_info 1961 * @alloc_fn: function to allocate percpu lpage, always called with lpage_size 1962 * @free_fn: function to free percpu memory, @size <= lpage_size 1963 * @map_fn: function to map percpu lpage, always called with lpage_size 1964 * 1965 * This allocator uses large page to build and map the first chunk. 1966 * Unlike other helpers, the caller should provide fully initialized 1967 * @ai. This can be done using pcpu_build_alloc_info(). This two 1968 * stage initialization is to allow arch code to evaluate the 1969 * parameters before committing to it. 1970 * 1971 * Large pages are allocated as directed by @unit_map and other 1972 * parameters and mapped to vmalloc space. Unused holes are returned 1973 * to the page allocator. Note that these holes end up being actively 1974 * mapped twice - once to the physical mapping and to the vmalloc area 1975 * for the first percpu chunk. Depending on architecture, this might 1976 * cause problem when changing page attributes of the returned area. 1977 * These double mapped areas can be detected using 1978 * pcpu_lpage_remapped(). 1979 * 1980 * RETURNS: 1981 * 0 on success, -errno on failure. 1982 */ 1983int __init pcpu_lpage_first_chunk(const struct pcpu_alloc_info *ai, 1984 pcpu_fc_alloc_fn_t alloc_fn, 1985 pcpu_fc_free_fn_t free_fn, 1986 pcpu_fc_map_fn_t map_fn) 1987{ 1988 static struct vm_struct vm; 1989 const size_t lpage_size = ai->atom_size; 1990 size_t chunk_size, map_size; 1991 unsigned int cpu; 1992 int i, j, unit, nr_units, rc; 1993 1994 nr_units = 0; 1995 for (i = 0; i < ai->nr_groups; i++) 1996 nr_units += ai->groups[i].nr_units; 1997 1998 chunk_size = ai->unit_size * nr_units; 1999 BUG_ON(chunk_size % lpage_size); 2000 2001 pcpul_size = ai->static_size + ai->reserved_size + ai->dyn_size; 2002 pcpul_lpage_size = lpage_size; 2003 pcpul_nr_lpages = chunk_size / lpage_size; 2004 2005 /* allocate pointer array and alloc large pages */ 2006 map_size = pcpul_nr_lpages * sizeof(pcpul_map[0]); 2007 pcpul_map = alloc_bootmem(map_size); 2008 2009 /* allocate all pages */ 2010 for (i = 0; i < pcpul_nr_lpages; i++) { 2011 size_t offset = i * lpage_size; 2012 int first_unit = offset / ai->unit_size; 2013 int last_unit = (offset + lpage_size - 1) / ai->unit_size; 2014 void *ptr; 2015 2016 /* find out which cpu is mapped to this unit */ 2017 for (unit = first_unit; unit <= last_unit; unit++) 2018 if (pcpul_unit_to_cpu(unit, ai, &cpu)) 2019 goto found; 2020 continue; 2021 found: 2022 ptr = alloc_fn(cpu, lpage_size, lpage_size); 2023 if (!ptr) { 2024 pr_warning("PERCPU: failed to allocate large page " 2025 "for cpu%u\n", cpu); 2026 goto enomem; 2027 } 2028 2029 pcpul_map[i].ptr = ptr; 2030 } 2031 2032 /* return unused holes */ 2033 for (unit = 0; unit < nr_units; unit++) { 2034 size_t start = unit * ai->unit_size; 2035 size_t end = start + ai->unit_size; 2036 size_t off, next; 2037 2038 /* don't free used part of occupied unit */ 2039 if (pcpul_unit_to_cpu(unit, ai, NULL)) 2040 start += pcpul_size; 2041 2042 /* unit can span more than one page, punch the holes */ 2043 for (off = start; off < end; off = next) { 2044 void *ptr = pcpul_map[off / lpage_size].ptr; 2045 next = min(roundup(off + 1, lpage_size), end); 2046 if (ptr) 2047 free_fn(ptr + off % lpage_size, next - off); 2048 } 2049 } 2050 2051 /* allocate address, map and copy */ 2052 vm.flags = VM_ALLOC; 2053 vm.size = chunk_size; 2054 vm_area_register_early(&vm, ai->unit_size); 2055 2056 for (i = 0; i < pcpul_nr_lpages; i++) { 2057 if (!pcpul_map[i].ptr) 2058 continue; 2059 pcpul_map[i].map_addr = vm.addr + i * lpage_size; 2060 map_fn(pcpul_map[i].ptr, lpage_size, pcpul_map[i].map_addr); 2061 } 2062 2063 for_each_possible_cpu(cpu) 2064 memcpy(vm.addr + pcpul_cpu_to_unit(cpu, ai) * ai->unit_size, 2065 __per_cpu_load, ai->static_size); 2066 2067 /* we're ready, commit */ 2068 pr_info("PERCPU: large pages @%p s%zu r%zu d%zu u%zu\n", 2069 vm.addr, ai->static_size, ai->reserved_size, ai->dyn_size, 2070 ai->unit_size); 2071 2072 rc = pcpu_setup_first_chunk(ai, vm.addr); 2073 2074 /* 2075 * Sort pcpul_map array for pcpu_lpage_remapped(). Unmapped 2076 * lpages are pushed to the end and trimmed. 2077 */ 2078 for (i = 0; i < pcpul_nr_lpages - 1; i++) 2079 for (j = i + 1; j < pcpul_nr_lpages; j++) { 2080 struct pcpul_ent tmp; 2081 2082 if (!pcpul_map[j].ptr) 2083 continue; 2084 if (pcpul_map[i].ptr && 2085 pcpul_map[i].ptr < pcpul_map[j].ptr) 2086 continue; 2087 2088 tmp = pcpul_map[i]; 2089 pcpul_map[i] = pcpul_map[j]; 2090 pcpul_map[j] = tmp; 2091 } 2092 2093 while (pcpul_nr_lpages && !pcpul_map[pcpul_nr_lpages - 1].ptr) 2094 pcpul_nr_lpages--; 2095 2096 return rc; 2097 2098enomem: 2099 for (i = 0; i < pcpul_nr_lpages; i++) 2100 if (pcpul_map[i].ptr) 2101 free_fn(pcpul_map[i].ptr, lpage_size); 2102 free_bootmem(__pa(pcpul_map), map_size); 2103 return -ENOMEM; 2104} 2105 2106/** 2107 * pcpu_lpage_remapped - determine whether a kaddr is in pcpul recycled area 2108 * @kaddr: the kernel address in question 2109 * 2110 * Determine whether @kaddr falls in the pcpul recycled area. This is 2111 * used by pageattr to detect VM aliases and break up the pcpu large 2112 * page mapping such that the same physical page is not mapped under 2113 * different attributes. 2114 * 2115 * The recycled area is always at the tail of a partially used large 2116 * page. 2117 * 2118 * RETURNS: 2119 * Address of corresponding remapped pcpu address if match is found; 2120 * otherwise, NULL. 2121 */ 2122void *pcpu_lpage_remapped(void *kaddr) 2123{ 2124 unsigned long lpage_mask = pcpul_lpage_size - 1; 2125 void *lpage_addr = (void *)((unsigned long)kaddr & ~lpage_mask); 2126 unsigned long offset = (unsigned long)kaddr & lpage_mask; 2127 int left = 0, right = pcpul_nr_lpages - 1; 2128 int pos; 2129 2130 /* pcpul in use at all? */ 2131 if (!pcpul_map) 2132 return NULL; 2133 2134 /* okay, perform binary search */ 2135 while (left <= right) { 2136 pos = (left + right) / 2; 2137 2138 if (pcpul_map[pos].ptr < lpage_addr) 2139 left = pos + 1; 2140 else if (pcpul_map[pos].ptr > lpage_addr) 2141 right = pos - 1; 2142 else 2143 return pcpul_map[pos].map_addr + offset; 2144 } 2145 2146 return NULL; 2147} 2148#endif /* CONFIG_NEED_PER_CPU_LPAGE_FIRST_CHUNK */ 2149 2150/* 2151 * Generic percpu area setup. 2152 * 2153 * The embedding helper is used because its behavior closely resembles 2154 * the original non-dynamic generic percpu area setup. This is 2155 * important because many archs have addressing restrictions and might 2156 * fail if the percpu area is located far away from the previous 2157 * location. As an added bonus, in non-NUMA cases, embedding is 2158 * generally a good idea TLB-wise because percpu area can piggy back 2159 * on the physical linear memory mapping which uses large page 2160 * mappings on applicable archs. 2161 */ 2162#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 2163unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 2164EXPORT_SYMBOL(__per_cpu_offset); 2165 2166void __init setup_per_cpu_areas(void) 2167{ 2168 unsigned long delta; 2169 unsigned int cpu; 2170 int rc; 2171 2172 /* 2173 * Always reserve area for module percpu variables. That's 2174 * what the legacy allocator did. 2175 */ 2176 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 2177 PERCPU_DYNAMIC_RESERVE); 2178 if (rc < 0) 2179 panic("Failed to initialized percpu areas."); 2180 2181 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 2182 for_each_possible_cpu(cpu) 2183 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 2184} 2185#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 2186