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 93 struct 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 }; 105 106 static int pcpu_unit_pages __read_mostly; 107 static int pcpu_unit_size __read_mostly; 108 static int pcpu_nr_units __read_mostly; 109 static int pcpu_atom_size __read_mostly; 110 static int pcpu_nr_slots __read_mostly; 111 static size_t pcpu_chunk_struct_size __read_mostly; 112 113 /* cpus with the lowest and highest unit numbers */ 114 static unsigned int pcpu_first_unit_cpu __read_mostly; 115 static unsigned int pcpu_last_unit_cpu __read_mostly; 116 117 /* the address of the first chunk which starts with the kernel static area */ 118 void *pcpu_base_addr __read_mostly; 119 EXPORT_SYMBOL_GPL(pcpu_base_addr); 120 121 static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */ 122 const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */ 123 124 /* group information, used for vm allocation */ 125 static int pcpu_nr_groups __read_mostly; 126 static const unsigned long *pcpu_group_offsets __read_mostly; 127 static 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 */ 134 static 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 */ 143 static struct pcpu_chunk *pcpu_reserved_chunk; 144 static 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 */ 167 static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ 168 static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ 169 170 static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ 171 172 /* reclaim work to release fully free chunks, scheduled from free path */ 173 static void pcpu_reclaim(struct work_struct *work); 174 static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); 175 176 static 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 } 181 182 static 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 } 188 189 static 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 } 196 197 static int pcpu_page_idx(unsigned int cpu, int page_idx) 198 { 199 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 200 } 201 202 static 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 } 208 209 static 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 */ 219 static 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 */ 225 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 226 { 227 return (struct pcpu_chunk *)page->index; 228 } 229 230 static 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 } 235 236 static 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 */ 272 static 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 */ 291 static 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 */ 312 static 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 */ 331 static 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[raw_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 */ 369 static 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 */ 434 static 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 */ 475 static 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 */ 564 static 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 */ 617 static 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 */ 654 static 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 */ 683 static 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 */ 718 static 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 } 725 726 static 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 */ 745 static 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 */ 781 static 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 } 788 789 static 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 */ 812 static 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; 836 837 err: 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 */ 859 static 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 */ 881 static 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 */ 934 static 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); 976 clear: 977 for_each_possible_cpu(cpu) 978 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); 979 return 0; 980 981 err_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); 986 err_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 } 991 992 static 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 1002 static 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 */ 1044 static void *pcpu_alloc(size_t size, size_t align, bool reserved) 1045 { 1046 static int warn_limit = 10; 1047 struct pcpu_chunk *chunk; 1048 const char *err; 1049 int slot, off; 1050 1051 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { 1052 WARN(true, "illegal size (%zu) or align (%zu) for " 1053 "percpu allocation\n", size, align); 1054 return NULL; 1055 } 1056 1057 mutex_lock(&pcpu_alloc_mutex); 1058 spin_lock_irq(&pcpu_lock); 1059 1060 /* serve reserved allocations from the reserved chunk if available */ 1061 if (reserved && pcpu_reserved_chunk) { 1062 chunk = pcpu_reserved_chunk; 1063 if (size > chunk->contig_hint || 1064 pcpu_extend_area_map(chunk) < 0) { 1065 err = "failed to extend area map of reserved chunk"; 1066 goto fail_unlock; 1067 } 1068 off = pcpu_alloc_area(chunk, size, align); 1069 if (off >= 0) 1070 goto area_found; 1071 err = "alloc from reserved chunk failed"; 1072 goto fail_unlock; 1073 } 1074 1075 restart: 1076 /* search through normal chunks */ 1077 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 1078 list_for_each_entry(chunk, &pcpu_slot[slot], list) { 1079 if (size > chunk->contig_hint) 1080 continue; 1081 1082 switch (pcpu_extend_area_map(chunk)) { 1083 case 0: 1084 break; 1085 case 1: 1086 goto restart; /* pcpu_lock dropped, restart */ 1087 default: 1088 err = "failed to extend area map"; 1089 goto fail_unlock; 1090 } 1091 1092 off = pcpu_alloc_area(chunk, size, align); 1093 if (off >= 0) 1094 goto area_found; 1095 } 1096 } 1097 1098 /* hmmm... no space left, create a new chunk */ 1099 spin_unlock_irq(&pcpu_lock); 1100 1101 chunk = alloc_pcpu_chunk(); 1102 if (!chunk) { 1103 err = "failed to allocate new chunk"; 1104 goto fail_unlock_mutex; 1105 } 1106 1107 spin_lock_irq(&pcpu_lock); 1108 pcpu_chunk_relocate(chunk, -1); 1109 goto restart; 1110 1111 area_found: 1112 spin_unlock_irq(&pcpu_lock); 1113 1114 /* populate, map and clear the area */ 1115 if (pcpu_populate_chunk(chunk, off, size)) { 1116 spin_lock_irq(&pcpu_lock); 1117 pcpu_free_area(chunk, off); 1118 err = "failed to populate"; 1119 goto fail_unlock; 1120 } 1121 1122 mutex_unlock(&pcpu_alloc_mutex); 1123 1124 /* return address relative to base address */ 1125 return __addr_to_pcpu_ptr(chunk->base_addr + off); 1126 1127 fail_unlock: 1128 spin_unlock_irq(&pcpu_lock); 1129 fail_unlock_mutex: 1130 mutex_unlock(&pcpu_alloc_mutex); 1131 if (warn_limit) { 1132 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, " 1133 "%s\n", size, align, err); 1134 dump_stack(); 1135 if (!--warn_limit) 1136 pr_info("PERCPU: limit reached, disable warning\n"); 1137 } 1138 return NULL; 1139 } 1140 1141 /** 1142 * __alloc_percpu - allocate dynamic percpu area 1143 * @size: size of area to allocate in bytes 1144 * @align: alignment of area (max PAGE_SIZE) 1145 * 1146 * Allocate percpu area of @size bytes aligned at @align. Might 1147 * sleep. Might trigger writeouts. 1148 * 1149 * CONTEXT: 1150 * Does GFP_KERNEL allocation. 1151 * 1152 * RETURNS: 1153 * Percpu pointer to the allocated area on success, NULL on failure. 1154 */ 1155 void *__alloc_percpu(size_t size, size_t align) 1156 { 1157 return pcpu_alloc(size, align, false); 1158 } 1159 EXPORT_SYMBOL_GPL(__alloc_percpu); 1160 1161 /** 1162 * __alloc_reserved_percpu - allocate reserved percpu area 1163 * @size: size of area to allocate in bytes 1164 * @align: alignment of area (max PAGE_SIZE) 1165 * 1166 * Allocate percpu area of @size bytes aligned at @align from reserved 1167 * percpu area if arch has set it up; otherwise, allocation is served 1168 * from the same dynamic area. Might sleep. Might trigger writeouts. 1169 * 1170 * CONTEXT: 1171 * Does GFP_KERNEL allocation. 1172 * 1173 * RETURNS: 1174 * Percpu pointer to the allocated area on success, NULL on failure. 1175 */ 1176 void *__alloc_reserved_percpu(size_t size, size_t align) 1177 { 1178 return pcpu_alloc(size, align, true); 1179 } 1180 1181 /** 1182 * pcpu_reclaim - reclaim fully free chunks, workqueue function 1183 * @work: unused 1184 * 1185 * Reclaim all fully free chunks except for the first one. 1186 * 1187 * CONTEXT: 1188 * workqueue context. 1189 */ 1190 static void pcpu_reclaim(struct work_struct *work) 1191 { 1192 LIST_HEAD(todo); 1193 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; 1194 struct pcpu_chunk *chunk, *next; 1195 1196 mutex_lock(&pcpu_alloc_mutex); 1197 spin_lock_irq(&pcpu_lock); 1198 1199 list_for_each_entry_safe(chunk, next, head, list) { 1200 WARN_ON(chunk->immutable); 1201 1202 /* spare the first one */ 1203 if (chunk == list_first_entry(head, struct pcpu_chunk, list)) 1204 continue; 1205 1206 list_move(&chunk->list, &todo); 1207 } 1208 1209 spin_unlock_irq(&pcpu_lock); 1210 1211 list_for_each_entry_safe(chunk, next, &todo, list) { 1212 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size); 1213 free_pcpu_chunk(chunk); 1214 } 1215 1216 mutex_unlock(&pcpu_alloc_mutex); 1217 } 1218 1219 /** 1220 * free_percpu - free percpu area 1221 * @ptr: pointer to area to free 1222 * 1223 * Free percpu area @ptr. 1224 * 1225 * CONTEXT: 1226 * Can be called from atomic context. 1227 */ 1228 void free_percpu(void *ptr) 1229 { 1230 void *addr = __pcpu_ptr_to_addr(ptr); 1231 struct pcpu_chunk *chunk; 1232 unsigned long flags; 1233 int off; 1234 1235 if (!ptr) 1236 return; 1237 1238 spin_lock_irqsave(&pcpu_lock, flags); 1239 1240 chunk = pcpu_chunk_addr_search(addr); 1241 off = addr - chunk->base_addr; 1242 1243 pcpu_free_area(chunk, off); 1244 1245 /* if there are more than one fully free chunks, wake up grim reaper */ 1246 if (chunk->free_size == pcpu_unit_size) { 1247 struct pcpu_chunk *pos; 1248 1249 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 1250 if (pos != chunk) { 1251 schedule_work(&pcpu_reclaim_work); 1252 break; 1253 } 1254 } 1255 1256 spin_unlock_irqrestore(&pcpu_lock, flags); 1257 } 1258 EXPORT_SYMBOL_GPL(free_percpu); 1259 1260 static inline size_t pcpu_calc_fc_sizes(size_t static_size, 1261 size_t reserved_size, 1262 ssize_t *dyn_sizep) 1263 { 1264 size_t size_sum; 1265 1266 size_sum = PFN_ALIGN(static_size + reserved_size + 1267 (*dyn_sizep >= 0 ? *dyn_sizep : 0)); 1268 if (*dyn_sizep != 0) 1269 *dyn_sizep = size_sum - static_size - reserved_size; 1270 1271 return size_sum; 1272 } 1273 1274 /** 1275 * pcpu_alloc_alloc_info - allocate percpu allocation info 1276 * @nr_groups: the number of groups 1277 * @nr_units: the number of units 1278 * 1279 * Allocate ai which is large enough for @nr_groups groups containing 1280 * @nr_units units. The returned ai's groups[0].cpu_map points to the 1281 * cpu_map array which is long enough for @nr_units and filled with 1282 * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1283 * pointer of other groups. 1284 * 1285 * RETURNS: 1286 * Pointer to the allocated pcpu_alloc_info on success, NULL on 1287 * failure. 1288 */ 1289 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 1290 int nr_units) 1291 { 1292 struct pcpu_alloc_info *ai; 1293 size_t base_size, ai_size; 1294 void *ptr; 1295 int unit; 1296 1297 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 1298 __alignof__(ai->groups[0].cpu_map[0])); 1299 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 1300 1301 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size)); 1302 if (!ptr) 1303 return NULL; 1304 ai = ptr; 1305 ptr += base_size; 1306 1307 ai->groups[0].cpu_map = ptr; 1308 1309 for (unit = 0; unit < nr_units; unit++) 1310 ai->groups[0].cpu_map[unit] = NR_CPUS; 1311 1312 ai->nr_groups = nr_groups; 1313 ai->__ai_size = PFN_ALIGN(ai_size); 1314 1315 return ai; 1316 } 1317 1318 /** 1319 * pcpu_free_alloc_info - free percpu allocation info 1320 * @ai: pcpu_alloc_info to free 1321 * 1322 * Free @ai which was allocated by pcpu_alloc_alloc_info(). 1323 */ 1324 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 1325 { 1326 free_bootmem(__pa(ai), ai->__ai_size); 1327 } 1328 1329 /** 1330 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 1331 * @reserved_size: the size of reserved percpu area in bytes 1332 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto 1333 * @atom_size: allocation atom size 1334 * @cpu_distance_fn: callback to determine distance between cpus, optional 1335 * 1336 * This function determines grouping of units, their mappings to cpus 1337 * and other parameters considering needed percpu size, allocation 1338 * atom size and distances between CPUs. 1339 * 1340 * Groups are always mutliples of atom size and CPUs which are of 1341 * LOCAL_DISTANCE both ways are grouped together and share space for 1342 * units in the same group. The returned configuration is guaranteed 1343 * to have CPUs on different nodes on different groups and >=75% usage 1344 * of allocated virtual address space. 1345 * 1346 * RETURNS: 1347 * On success, pointer to the new allocation_info is returned. On 1348 * failure, ERR_PTR value is returned. 1349 */ 1350 struct pcpu_alloc_info * __init pcpu_build_alloc_info( 1351 size_t reserved_size, ssize_t dyn_size, 1352 size_t atom_size, 1353 pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 1354 { 1355 static int group_map[NR_CPUS] __initdata; 1356 static int group_cnt[NR_CPUS] __initdata; 1357 const size_t static_size = __per_cpu_end - __per_cpu_start; 1358 int group_cnt_max = 0, nr_groups = 1, nr_units = 0; 1359 size_t size_sum, min_unit_size, alloc_size; 1360 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 1361 int last_allocs, group, unit; 1362 unsigned int cpu, tcpu; 1363 struct pcpu_alloc_info *ai; 1364 unsigned int *cpu_map; 1365 1366 /* this function may be called multiple times */ 1367 memset(group_map, 0, sizeof(group_map)); 1368 memset(group_cnt, 0, sizeof(group_map)); 1369 1370 /* 1371 * Determine min_unit_size, alloc_size and max_upa such that 1372 * alloc_size is multiple of atom_size and is the smallest 1373 * which can accomodate 4k aligned segments which are equal to 1374 * or larger than min_unit_size. 1375 */ 1376 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size); 1377 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 1378 1379 alloc_size = roundup(min_unit_size, atom_size); 1380 upa = alloc_size / min_unit_size; 1381 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1382 upa--; 1383 max_upa = upa; 1384 1385 /* group cpus according to their proximity */ 1386 for_each_possible_cpu(cpu) { 1387 group = 0; 1388 next_group: 1389 for_each_possible_cpu(tcpu) { 1390 if (cpu == tcpu) 1391 break; 1392 if (group_map[tcpu] == group && cpu_distance_fn && 1393 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 1394 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 1395 group++; 1396 nr_groups = max(nr_groups, group + 1); 1397 goto next_group; 1398 } 1399 } 1400 group_map[cpu] = group; 1401 group_cnt[group]++; 1402 group_cnt_max = max(group_cnt_max, group_cnt[group]); 1403 } 1404 1405 /* 1406 * Expand unit size until address space usage goes over 75% 1407 * and then as much as possible without using more address 1408 * space. 1409 */ 1410 last_allocs = INT_MAX; 1411 for (upa = max_upa; upa; upa--) { 1412 int allocs = 0, wasted = 0; 1413 1414 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1415 continue; 1416 1417 for (group = 0; group < nr_groups; group++) { 1418 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 1419 allocs += this_allocs; 1420 wasted += this_allocs * upa - group_cnt[group]; 1421 } 1422 1423 /* 1424 * Don't accept if wastage is over 25%. The 1425 * greater-than comparison ensures upa==1 always 1426 * passes the following check. 1427 */ 1428 if (wasted > num_possible_cpus() / 3) 1429 continue; 1430 1431 /* and then don't consume more memory */ 1432 if (allocs > last_allocs) 1433 break; 1434 last_allocs = allocs; 1435 best_upa = upa; 1436 } 1437 upa = best_upa; 1438 1439 /* allocate and fill alloc_info */ 1440 for (group = 0; group < nr_groups; group++) 1441 nr_units += roundup(group_cnt[group], upa); 1442 1443 ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 1444 if (!ai) 1445 return ERR_PTR(-ENOMEM); 1446 cpu_map = ai->groups[0].cpu_map; 1447 1448 for (group = 0; group < nr_groups; group++) { 1449 ai->groups[group].cpu_map = cpu_map; 1450 cpu_map += roundup(group_cnt[group], upa); 1451 } 1452 1453 ai->static_size = static_size; 1454 ai->reserved_size = reserved_size; 1455 ai->dyn_size = dyn_size; 1456 ai->unit_size = alloc_size / upa; 1457 ai->atom_size = atom_size; 1458 ai->alloc_size = alloc_size; 1459 1460 for (group = 0, unit = 0; group_cnt[group]; group++) { 1461 struct pcpu_group_info *gi = &ai->groups[group]; 1462 1463 /* 1464 * Initialize base_offset as if all groups are located 1465 * back-to-back. The caller should update this to 1466 * reflect actual allocation. 1467 */ 1468 gi->base_offset = unit * ai->unit_size; 1469 1470 for_each_possible_cpu(cpu) 1471 if (group_map[cpu] == group) 1472 gi->cpu_map[gi->nr_units++] = cpu; 1473 gi->nr_units = roundup(gi->nr_units, upa); 1474 unit += gi->nr_units; 1475 } 1476 BUG_ON(unit != nr_units); 1477 1478 return ai; 1479 } 1480 1481 /** 1482 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 1483 * @lvl: loglevel 1484 * @ai: allocation info to dump 1485 * 1486 * Print out information about @ai using loglevel @lvl. 1487 */ 1488 static void pcpu_dump_alloc_info(const char *lvl, 1489 const struct pcpu_alloc_info *ai) 1490 { 1491 int group_width = 1, cpu_width = 1, width; 1492 char empty_str[] = "--------"; 1493 int alloc = 0, alloc_end = 0; 1494 int group, v; 1495 int upa, apl; /* units per alloc, allocs per line */ 1496 1497 v = ai->nr_groups; 1498 while (v /= 10) 1499 group_width++; 1500 1501 v = num_possible_cpus(); 1502 while (v /= 10) 1503 cpu_width++; 1504 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 1505 1506 upa = ai->alloc_size / ai->unit_size; 1507 width = upa * (cpu_width + 1) + group_width + 3; 1508 apl = rounddown_pow_of_two(max(60 / width, 1)); 1509 1510 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 1511 lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 1512 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 1513 1514 for (group = 0; group < ai->nr_groups; group++) { 1515 const struct pcpu_group_info *gi = &ai->groups[group]; 1516 int unit = 0, unit_end = 0; 1517 1518 BUG_ON(gi->nr_units % upa); 1519 for (alloc_end += gi->nr_units / upa; 1520 alloc < alloc_end; alloc++) { 1521 if (!(alloc % apl)) { 1522 printk("\n"); 1523 printk("%spcpu-alloc: ", lvl); 1524 } 1525 printk("[%0*d] ", group_width, group); 1526 1527 for (unit_end += upa; unit < unit_end; unit++) 1528 if (gi->cpu_map[unit] != NR_CPUS) 1529 printk("%0*d ", cpu_width, 1530 gi->cpu_map[unit]); 1531 else 1532 printk("%s ", empty_str); 1533 } 1534 } 1535 printk("\n"); 1536 } 1537 1538 /** 1539 * pcpu_setup_first_chunk - initialize the first percpu chunk 1540 * @ai: pcpu_alloc_info describing how to percpu area is shaped 1541 * @base_addr: mapped address 1542 * 1543 * Initialize the first percpu chunk which contains the kernel static 1544 * perpcu area. This function is to be called from arch percpu area 1545 * setup path. 1546 * 1547 * @ai contains all information necessary to initialize the first 1548 * chunk and prime the dynamic percpu allocator. 1549 * 1550 * @ai->static_size is the size of static percpu area. 1551 * 1552 * @ai->reserved_size, if non-zero, specifies the amount of bytes to 1553 * reserve after the static area in the first chunk. This reserves 1554 * the first chunk such that it's available only through reserved 1555 * percpu allocation. This is primarily used to serve module percpu 1556 * static areas on architectures where the addressing model has 1557 * limited offset range for symbol relocations to guarantee module 1558 * percpu symbols fall inside the relocatable range. 1559 * 1560 * @ai->dyn_size determines the number of bytes available for dynamic 1561 * allocation in the first chunk. The area between @ai->static_size + 1562 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 1563 * 1564 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 1565 * and equal to or larger than @ai->static_size + @ai->reserved_size + 1566 * @ai->dyn_size. 1567 * 1568 * @ai->atom_size is the allocation atom size and used as alignment 1569 * for vm areas. 1570 * 1571 * @ai->alloc_size is the allocation size and always multiple of 1572 * @ai->atom_size. This is larger than @ai->atom_size if 1573 * @ai->unit_size is larger than @ai->atom_size. 1574 * 1575 * @ai->nr_groups and @ai->groups describe virtual memory layout of 1576 * percpu areas. Units which should be colocated are put into the 1577 * same group. Dynamic VM areas will be allocated according to these 1578 * groupings. If @ai->nr_groups is zero, a single group containing 1579 * all units is assumed. 1580 * 1581 * The caller should have mapped the first chunk at @base_addr and 1582 * copied static data to each unit. 1583 * 1584 * If the first chunk ends up with both reserved and dynamic areas, it 1585 * is served by two chunks - one to serve the core static and reserved 1586 * areas and the other for the dynamic area. They share the same vm 1587 * and page map but uses different area allocation map to stay away 1588 * from each other. The latter chunk is circulated in the chunk slots 1589 * and available for dynamic allocation like any other chunks. 1590 * 1591 * RETURNS: 1592 * 0 on success, -errno on failure. 1593 */ 1594 int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 1595 void *base_addr) 1596 { 1597 static char cpus_buf[4096] __initdata; 1598 static int smap[2], dmap[2]; 1599 size_t dyn_size = ai->dyn_size; 1600 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; 1601 struct pcpu_chunk *schunk, *dchunk = NULL; 1602 unsigned long *group_offsets; 1603 size_t *group_sizes; 1604 unsigned long *unit_off; 1605 unsigned int cpu; 1606 int *unit_map; 1607 int group, unit, i; 1608 1609 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask); 1610 1611 #define PCPU_SETUP_BUG_ON(cond) do { \ 1612 if (unlikely(cond)) { \ 1613 pr_emerg("PERCPU: failed to initialize, %s", #cond); \ 1614 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \ 1615 pcpu_dump_alloc_info(KERN_EMERG, ai); \ 1616 BUG(); \ 1617 } \ 1618 } while (0) 1619 1620 /* sanity checks */ 1621 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || 1622 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); 1623 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 1624 PCPU_SETUP_BUG_ON(!ai->static_size); 1625 PCPU_SETUP_BUG_ON(!base_addr); 1626 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 1627 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK); 1628 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 1629 1630 /* process group information and build config tables accordingly */ 1631 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0])); 1632 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0])); 1633 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0])); 1634 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0])); 1635 1636 for (cpu = 0; cpu < nr_cpu_ids; cpu++) 1637 unit_map[cpu] = UINT_MAX; 1638 pcpu_first_unit_cpu = NR_CPUS; 1639 1640 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 1641 const struct pcpu_group_info *gi = &ai->groups[group]; 1642 1643 group_offsets[group] = gi->base_offset; 1644 group_sizes[group] = gi->nr_units * ai->unit_size; 1645 1646 for (i = 0; i < gi->nr_units; i++) { 1647 cpu = gi->cpu_map[i]; 1648 if (cpu == NR_CPUS) 1649 continue; 1650 1651 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids); 1652 PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 1653 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 1654 1655 unit_map[cpu] = unit + i; 1656 unit_off[cpu] = gi->base_offset + i * ai->unit_size; 1657 1658 if (pcpu_first_unit_cpu == NR_CPUS) 1659 pcpu_first_unit_cpu = cpu; 1660 } 1661 } 1662 pcpu_last_unit_cpu = cpu; 1663 pcpu_nr_units = unit; 1664 1665 for_each_possible_cpu(cpu) 1666 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 1667 1668 /* we're done parsing the input, undefine BUG macro and dump config */ 1669 #undef PCPU_SETUP_BUG_ON 1670 pcpu_dump_alloc_info(KERN_INFO, ai); 1671 1672 pcpu_nr_groups = ai->nr_groups; 1673 pcpu_group_offsets = group_offsets; 1674 pcpu_group_sizes = group_sizes; 1675 pcpu_unit_map = unit_map; 1676 pcpu_unit_offsets = unit_off; 1677 1678 /* determine basic parameters */ 1679 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 1680 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 1681 pcpu_atom_size = ai->atom_size; 1682 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 1683 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 1684 1685 /* 1686 * Allocate chunk slots. The additional last slot is for 1687 * empty chunks. 1688 */ 1689 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 1690 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); 1691 for (i = 0; i < pcpu_nr_slots; i++) 1692 INIT_LIST_HEAD(&pcpu_slot[i]); 1693 1694 /* 1695 * Initialize static chunk. If reserved_size is zero, the 1696 * static chunk covers static area + dynamic allocation area 1697 * in the first chunk. If reserved_size is not zero, it 1698 * covers static area + reserved area (mostly used for module 1699 * static percpu allocation). 1700 */ 1701 schunk = alloc_bootmem(pcpu_chunk_struct_size); 1702 INIT_LIST_HEAD(&schunk->list); 1703 schunk->base_addr = base_addr; 1704 schunk->map = smap; 1705 schunk->map_alloc = ARRAY_SIZE(smap); 1706 schunk->immutable = true; 1707 bitmap_fill(schunk->populated, pcpu_unit_pages); 1708 1709 if (ai->reserved_size) { 1710 schunk->free_size = ai->reserved_size; 1711 pcpu_reserved_chunk = schunk; 1712 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; 1713 } else { 1714 schunk->free_size = dyn_size; 1715 dyn_size = 0; /* dynamic area covered */ 1716 } 1717 schunk->contig_hint = schunk->free_size; 1718 1719 schunk->map[schunk->map_used++] = -ai->static_size; 1720 if (schunk->free_size) 1721 schunk->map[schunk->map_used++] = schunk->free_size; 1722 1723 /* init dynamic chunk if necessary */ 1724 if (dyn_size) { 1725 dchunk = alloc_bootmem(pcpu_chunk_struct_size); 1726 INIT_LIST_HEAD(&dchunk->list); 1727 dchunk->base_addr = base_addr; 1728 dchunk->map = dmap; 1729 dchunk->map_alloc = ARRAY_SIZE(dmap); 1730 dchunk->immutable = true; 1731 bitmap_fill(dchunk->populated, pcpu_unit_pages); 1732 1733 dchunk->contig_hint = dchunk->free_size = dyn_size; 1734 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; 1735 dchunk->map[dchunk->map_used++] = dchunk->free_size; 1736 } 1737 1738 /* link the first chunk in */ 1739 pcpu_first_chunk = dchunk ?: schunk; 1740 pcpu_chunk_relocate(pcpu_first_chunk, -1); 1741 1742 /* we're done */ 1743 pcpu_base_addr = base_addr; 1744 return 0; 1745 } 1746 1747 const char *pcpu_fc_names[PCPU_FC_NR] __initdata = { 1748 [PCPU_FC_AUTO] = "auto", 1749 [PCPU_FC_EMBED] = "embed", 1750 [PCPU_FC_PAGE] = "page", 1751 }; 1752 1753 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 1754 1755 static int __init percpu_alloc_setup(char *str) 1756 { 1757 if (0) 1758 /* nada */; 1759 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 1760 else if (!strcmp(str, "embed")) 1761 pcpu_chosen_fc = PCPU_FC_EMBED; 1762 #endif 1763 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1764 else if (!strcmp(str, "page")) 1765 pcpu_chosen_fc = PCPU_FC_PAGE; 1766 #endif 1767 else 1768 pr_warning("PERCPU: unknown allocator %s specified\n", str); 1769 1770 return 0; 1771 } 1772 early_param("percpu_alloc", percpu_alloc_setup); 1773 1774 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 1775 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 1776 /** 1777 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 1778 * @reserved_size: the size of reserved percpu area in bytes 1779 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto 1780 * @atom_size: allocation atom size 1781 * @cpu_distance_fn: callback to determine distance between cpus, optional 1782 * @alloc_fn: function to allocate percpu page 1783 * @free_fn: funtion to free percpu page 1784 * 1785 * This is a helper to ease setting up embedded first percpu chunk and 1786 * can be called where pcpu_setup_first_chunk() is expected. 1787 * 1788 * If this function is used to setup the first chunk, it is allocated 1789 * by calling @alloc_fn and used as-is without being mapped into 1790 * vmalloc area. Allocations are always whole multiples of @atom_size 1791 * aligned to @atom_size. 1792 * 1793 * This enables the first chunk to piggy back on the linear physical 1794 * mapping which often uses larger page size. Please note that this 1795 * can result in very sparse cpu->unit mapping on NUMA machines thus 1796 * requiring large vmalloc address space. Don't use this allocator if 1797 * vmalloc space is not orders of magnitude larger than distances 1798 * between node memory addresses (ie. 32bit NUMA machines). 1799 * 1800 * When @dyn_size is positive, dynamic area might be larger than 1801 * specified to fill page alignment. When @dyn_size is auto, 1802 * @dyn_size is just big enough to fill page alignment after static 1803 * and reserved areas. 1804 * 1805 * If the needed size is smaller than the minimum or specified unit 1806 * size, the leftover is returned using @free_fn. 1807 * 1808 * RETURNS: 1809 * 0 on success, -errno on failure. 1810 */ 1811 int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size, 1812 size_t atom_size, 1813 pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 1814 pcpu_fc_alloc_fn_t alloc_fn, 1815 pcpu_fc_free_fn_t free_fn) 1816 { 1817 void *base = (void *)ULONG_MAX; 1818 void **areas = NULL; 1819 struct pcpu_alloc_info *ai; 1820 size_t size_sum, areas_size, max_distance; 1821 int group, i, rc; 1822 1823 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 1824 cpu_distance_fn); 1825 if (IS_ERR(ai)) 1826 return PTR_ERR(ai); 1827 1828 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 1829 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 1830 1831 areas = alloc_bootmem_nopanic(areas_size); 1832 if (!areas) { 1833 rc = -ENOMEM; 1834 goto out_free; 1835 } 1836 1837 /* allocate, copy and determine base address */ 1838 for (group = 0; group < ai->nr_groups; group++) { 1839 struct pcpu_group_info *gi = &ai->groups[group]; 1840 unsigned int cpu = NR_CPUS; 1841 void *ptr; 1842 1843 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 1844 cpu = gi->cpu_map[i]; 1845 BUG_ON(cpu == NR_CPUS); 1846 1847 /* allocate space for the whole group */ 1848 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 1849 if (!ptr) { 1850 rc = -ENOMEM; 1851 goto out_free_areas; 1852 } 1853 areas[group] = ptr; 1854 1855 base = min(ptr, base); 1856 1857 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 1858 if (gi->cpu_map[i] == NR_CPUS) { 1859 /* unused unit, free whole */ 1860 free_fn(ptr, ai->unit_size); 1861 continue; 1862 } 1863 /* copy and return the unused part */ 1864 memcpy(ptr, __per_cpu_load, ai->static_size); 1865 free_fn(ptr + size_sum, ai->unit_size - size_sum); 1866 } 1867 } 1868 1869 /* base address is now known, determine group base offsets */ 1870 max_distance = 0; 1871 for (group = 0; group < ai->nr_groups; group++) { 1872 ai->groups[group].base_offset = areas[group] - base; 1873 max_distance = max(max_distance, ai->groups[group].base_offset); 1874 } 1875 max_distance += ai->unit_size; 1876 1877 /* warn if maximum distance is further than 75% of vmalloc space */ 1878 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) { 1879 pr_warning("PERCPU: max_distance=0x%lx too large for vmalloc " 1880 "space 0x%lx\n", 1881 max_distance, VMALLOC_END - VMALLOC_START); 1882 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1883 /* and fail if we have fallback */ 1884 rc = -EINVAL; 1885 goto out_free; 1886 #endif 1887 } 1888 1889 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 1890 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 1891 ai->dyn_size, ai->unit_size); 1892 1893 rc = pcpu_setup_first_chunk(ai, base); 1894 goto out_free; 1895 1896 out_free_areas: 1897 for (group = 0; group < ai->nr_groups; group++) 1898 free_fn(areas[group], 1899 ai->groups[group].nr_units * ai->unit_size); 1900 out_free: 1901 pcpu_free_alloc_info(ai); 1902 if (areas) 1903 free_bootmem(__pa(areas), areas_size); 1904 return rc; 1905 } 1906 #endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK || 1907 !CONFIG_HAVE_SETUP_PER_CPU_AREA */ 1908 1909 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1910 /** 1911 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 1912 * @reserved_size: the size of reserved percpu area in bytes 1913 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 1914 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE 1915 * @populate_pte_fn: function to populate pte 1916 * 1917 * This is a helper to ease setting up page-remapped first percpu 1918 * chunk and can be called where pcpu_setup_first_chunk() is expected. 1919 * 1920 * This is the basic allocator. Static percpu area is allocated 1921 * page-by-page into vmalloc area. 1922 * 1923 * RETURNS: 1924 * 0 on success, -errno on failure. 1925 */ 1926 int __init pcpu_page_first_chunk(size_t reserved_size, 1927 pcpu_fc_alloc_fn_t alloc_fn, 1928 pcpu_fc_free_fn_t free_fn, 1929 pcpu_fc_populate_pte_fn_t populate_pte_fn) 1930 { 1931 static struct vm_struct vm; 1932 struct pcpu_alloc_info *ai; 1933 char psize_str[16]; 1934 int unit_pages; 1935 size_t pages_size; 1936 struct page **pages; 1937 int unit, i, j, rc; 1938 1939 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 1940 1941 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL); 1942 if (IS_ERR(ai)) 1943 return PTR_ERR(ai); 1944 BUG_ON(ai->nr_groups != 1); 1945 BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); 1946 1947 unit_pages = ai->unit_size >> PAGE_SHIFT; 1948 1949 /* unaligned allocations can't be freed, round up to page size */ 1950 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 1951 sizeof(pages[0])); 1952 pages = alloc_bootmem(pages_size); 1953 1954 /* allocate pages */ 1955 j = 0; 1956 for (unit = 0; unit < num_possible_cpus(); unit++) 1957 for (i = 0; i < unit_pages; i++) { 1958 unsigned int cpu = ai->groups[0].cpu_map[unit]; 1959 void *ptr; 1960 1961 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 1962 if (!ptr) { 1963 pr_warning("PERCPU: failed to allocate %s page " 1964 "for cpu%u\n", psize_str, cpu); 1965 goto enomem; 1966 } 1967 pages[j++] = virt_to_page(ptr); 1968 } 1969 1970 /* allocate vm area, map the pages and copy static data */ 1971 vm.flags = VM_ALLOC; 1972 vm.size = num_possible_cpus() * ai->unit_size; 1973 vm_area_register_early(&vm, PAGE_SIZE); 1974 1975 for (unit = 0; unit < num_possible_cpus(); unit++) { 1976 unsigned long unit_addr = 1977 (unsigned long)vm.addr + unit * ai->unit_size; 1978 1979 for (i = 0; i < unit_pages; i++) 1980 populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 1981 1982 /* pte already populated, the following shouldn't fail */ 1983 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 1984 unit_pages); 1985 if (rc < 0) 1986 panic("failed to map percpu area, err=%d\n", rc); 1987 1988 /* 1989 * FIXME: Archs with virtual cache should flush local 1990 * cache for the linear mapping here - something 1991 * equivalent to flush_cache_vmap() on the local cpu. 1992 * flush_cache_vmap() can't be used as most supporting 1993 * data structures are not set up yet. 1994 */ 1995 1996 /* copy static data */ 1997 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 1998 } 1999 2000 /* we're ready, commit */ 2001 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n", 2002 unit_pages, psize_str, vm.addr, ai->static_size, 2003 ai->reserved_size, ai->dyn_size); 2004 2005 rc = pcpu_setup_first_chunk(ai, vm.addr); 2006 goto out_free_ar; 2007 2008 enomem: 2009 while (--j >= 0) 2010 free_fn(page_address(pages[j]), PAGE_SIZE); 2011 rc = -ENOMEM; 2012 out_free_ar: 2013 free_bootmem(__pa(pages), pages_size); 2014 pcpu_free_alloc_info(ai); 2015 return rc; 2016 } 2017 #endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */ 2018 2019 /* 2020 * Generic percpu area setup. 2021 * 2022 * The embedding helper is used because its behavior closely resembles 2023 * the original non-dynamic generic percpu area setup. This is 2024 * important because many archs have addressing restrictions and might 2025 * fail if the percpu area is located far away from the previous 2026 * location. As an added bonus, in non-NUMA cases, embedding is 2027 * generally a good idea TLB-wise because percpu area can piggy back 2028 * on the physical linear memory mapping which uses large page 2029 * mappings on applicable archs. 2030 */ 2031 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 2032 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 2033 EXPORT_SYMBOL(__per_cpu_offset); 2034 2035 static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 2036 size_t align) 2037 { 2038 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS)); 2039 } 2040 2041 static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 2042 { 2043 free_bootmem(__pa(ptr), size); 2044 } 2045 2046 void __init setup_per_cpu_areas(void) 2047 { 2048 unsigned long delta; 2049 unsigned int cpu; 2050 int rc; 2051 2052 /* 2053 * Always reserve area for module percpu variables. That's 2054 * what the legacy allocator did. 2055 */ 2056 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 2057 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 2058 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 2059 if (rc < 0) 2060 panic("Failed to initialized percpu areas."); 2061 2062 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 2063 for_each_possible_cpu(cpu) 2064 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 2065 } 2066 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 2067