1 /* 2 * 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. Each chunk is 11 * consisted of boot-time determined number of units and the first 12 * 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. 17 * 18 * c0 c1 c2 19 * ------------------- ------------------- ------------ 20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u 21 * ------------------- ...... ------------------- .... ------------ 22 * 23 * Allocation is done in offset-size areas of single unit space. Ie, 24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0, 25 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to 26 * cpus. On NUMA, the mapping can be non-linear and even sparse. 27 * Percpu access can be done by configuring percpu base registers 28 * according to cpu to unit mapping and pcpu_unit_size. 29 * 30 * There are usually many small percpu allocations many of them being 31 * as small as 4 bytes. The allocator organizes chunks into lists 32 * according to free size and tries to allocate from the fullest one. 33 * Each chunk keeps the maximum contiguous area size hint which is 34 * guaranteed to be eqaul to or larger than the maximum contiguous 35 * area in the chunk. This helps the allocator not to iterate the 36 * chunk maps unnecessarily. 37 * 38 * Allocation state in each chunk is kept using an array of integers 39 * on chunk->map. A positive value in the map represents a free 40 * region and negative allocated. Allocation inside a chunk is done 41 * by scanning this map sequentially and serving the first matching 42 * entry. This is mostly copied from the percpu_modalloc() allocator. 43 * Chunks can be determined from the address using the index field 44 * in the page struct. The index field contains a pointer to the chunk. 45 * 46 * To use this allocator, arch code should do the followings. 47 * 48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate 49 * regular address to percpu pointer and back if they need to be 50 * different from the default 51 * 52 * - use pcpu_setup_first_chunk() during percpu area initialization to 53 * setup the first chunk containing the kernel static percpu area 54 */ 55 56 #include <linux/bitmap.h> 57 #include <linux/bootmem.h> 58 #include <linux/err.h> 59 #include <linux/list.h> 60 #include <linux/log2.h> 61 #include <linux/mm.h> 62 #include <linux/module.h> 63 #include <linux/mutex.h> 64 #include <linux/percpu.h> 65 #include <linux/pfn.h> 66 #include <linux/slab.h> 67 #include <linux/spinlock.h> 68 #include <linux/vmalloc.h> 69 #include <linux/workqueue.h> 70 71 #include <asm/cacheflush.h> 72 #include <asm/sections.h> 73 #include <asm/tlbflush.h> 74 #include <asm/io.h> 75 76 #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ 77 #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ 78 79 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ 80 #ifndef __addr_to_pcpu_ptr 81 #define __addr_to_pcpu_ptr(addr) \ 82 (void __percpu *)((unsigned long)(addr) - \ 83 (unsigned long)pcpu_base_addr + \ 84 (unsigned long)__per_cpu_start) 85 #endif 86 #ifndef __pcpu_ptr_to_addr 87 #define __pcpu_ptr_to_addr(ptr) \ 88 (void __force *)((unsigned long)(ptr) + \ 89 (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 void *data; /* chunk data */ 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. In 157 * general, percpu memory can't be allocated with irq off but 158 * irqsave/restore are still used in alloc path so that it can be used 159 * from early init path - sched_init() specifically. 160 * 161 * Free path accesses and alters only the index data structures, so it 162 * can be safely called from atomic context. When memory needs to be 163 * returned to the system, free path schedules reclaim_work which 164 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be 165 * reclaimed, release both locks and frees the chunks. Note that it's 166 * necessary to grab both locks to remove a chunk from circulation as 167 * allocation path might be referencing the chunk with only 168 * pcpu_alloc_mutex locked. 169 */ 170 static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ 171 static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ 172 173 static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ 174 175 /* reclaim work to release fully free chunks, scheduled from free path */ 176 static void pcpu_reclaim(struct work_struct *work); 177 static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); 178 179 static bool pcpu_addr_in_first_chunk(void *addr) 180 { 181 void *first_start = pcpu_first_chunk->base_addr; 182 183 return addr >= first_start && addr < first_start + pcpu_unit_size; 184 } 185 186 static bool pcpu_addr_in_reserved_chunk(void *addr) 187 { 188 void *first_start = pcpu_first_chunk->base_addr; 189 190 return addr >= first_start && 191 addr < first_start + pcpu_reserved_chunk_limit; 192 } 193 194 static int __pcpu_size_to_slot(int size) 195 { 196 int highbit = fls(size); /* size is in bytes */ 197 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 198 } 199 200 static int pcpu_size_to_slot(int size) 201 { 202 if (size == pcpu_unit_size) 203 return pcpu_nr_slots - 1; 204 return __pcpu_size_to_slot(size); 205 } 206 207 static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 208 { 209 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) 210 return 0; 211 212 return pcpu_size_to_slot(chunk->free_size); 213 } 214 215 /* set the pointer to a chunk in a page struct */ 216 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 217 { 218 page->index = (unsigned long)pcpu; 219 } 220 221 /* obtain pointer to a chunk from a page struct */ 222 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 223 { 224 return (struct pcpu_chunk *)page->index; 225 } 226 227 static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) 228 { 229 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 230 } 231 232 static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 233 unsigned int cpu, int page_idx) 234 { 235 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] + 236 (page_idx << PAGE_SHIFT); 237 } 238 239 static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk, 240 int *rs, int *re, int end) 241 { 242 *rs = find_next_zero_bit(chunk->populated, end, *rs); 243 *re = find_next_bit(chunk->populated, end, *rs + 1); 244 } 245 246 static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk, 247 int *rs, int *re, int end) 248 { 249 *rs = find_next_bit(chunk->populated, end, *rs); 250 *re = find_next_zero_bit(chunk->populated, end, *rs + 1); 251 } 252 253 /* 254 * (Un)populated page region iterators. Iterate over (un)populated 255 * page regions betwen @start and @end in @chunk. @rs and @re should 256 * be integer variables and will be set to start and end page index of 257 * the current region. 258 */ 259 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ 260 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ 261 (rs) < (re); \ 262 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) 263 264 #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ 265 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ 266 (rs) < (re); \ 267 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) 268 269 /** 270 * pcpu_mem_alloc - allocate memory 271 * @size: bytes to allocate 272 * 273 * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 274 * kzalloc() is used; otherwise, vmalloc() is used. The returned 275 * memory is always zeroed. 276 * 277 * CONTEXT: 278 * Does GFP_KERNEL allocation. 279 * 280 * RETURNS: 281 * Pointer to the allocated area on success, NULL on failure. 282 */ 283 static void *pcpu_mem_alloc(size_t size) 284 { 285 if (WARN_ON_ONCE(!slab_is_available())) 286 return NULL; 287 288 if (size <= PAGE_SIZE) 289 return kzalloc(size, GFP_KERNEL); 290 else { 291 void *ptr = vmalloc(size); 292 if (ptr) 293 memset(ptr, 0, size); 294 return ptr; 295 } 296 } 297 298 /** 299 * pcpu_mem_free - free memory 300 * @ptr: memory to free 301 * @size: size of the area 302 * 303 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). 304 */ 305 static void pcpu_mem_free(void *ptr, size_t size) 306 { 307 if (size <= PAGE_SIZE) 308 kfree(ptr); 309 else 310 vfree(ptr); 311 } 312 313 /** 314 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 315 * @chunk: chunk of interest 316 * @oslot: the previous slot it was on 317 * 318 * This function is called after an allocation or free changed @chunk. 319 * New slot according to the changed state is determined and @chunk is 320 * moved to the slot. Note that the reserved chunk is never put on 321 * chunk slots. 322 * 323 * CONTEXT: 324 * pcpu_lock. 325 */ 326 static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 327 { 328 int nslot = pcpu_chunk_slot(chunk); 329 330 if (chunk != pcpu_reserved_chunk && oslot != nslot) { 331 if (oslot < nslot) 332 list_move(&chunk->list, &pcpu_slot[nslot]); 333 else 334 list_move_tail(&chunk->list, &pcpu_slot[nslot]); 335 } 336 } 337 338 /** 339 * pcpu_need_to_extend - determine whether chunk area map needs to be extended 340 * @chunk: chunk of interest 341 * 342 * Determine whether area map of @chunk needs to be extended to 343 * accomodate a new allocation. 344 * 345 * CONTEXT: 346 * pcpu_lock. 347 * 348 * RETURNS: 349 * New target map allocation length if extension is necessary, 0 350 * otherwise. 351 */ 352 static int pcpu_need_to_extend(struct pcpu_chunk *chunk) 353 { 354 int new_alloc; 355 356 if (chunk->map_alloc >= chunk->map_used + 2) 357 return 0; 358 359 new_alloc = PCPU_DFL_MAP_ALLOC; 360 while (new_alloc < chunk->map_used + 2) 361 new_alloc *= 2; 362 363 return new_alloc; 364 } 365 366 /** 367 * pcpu_extend_area_map - extend area map of a chunk 368 * @chunk: chunk of interest 369 * @new_alloc: new target allocation length of the area map 370 * 371 * Extend area map of @chunk to have @new_alloc entries. 372 * 373 * CONTEXT: 374 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock. 375 * 376 * RETURNS: 377 * 0 on success, -errno on failure. 378 */ 379 static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc) 380 { 381 int *old = NULL, *new = NULL; 382 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]); 383 unsigned long flags; 384 385 new = pcpu_mem_alloc(new_size); 386 if (!new) 387 return -ENOMEM; 388 389 /* acquire pcpu_lock and switch to new area map */ 390 spin_lock_irqsave(&pcpu_lock, flags); 391 392 if (new_alloc <= chunk->map_alloc) 393 goto out_unlock; 394 395 old_size = chunk->map_alloc * sizeof(chunk->map[0]); 396 memcpy(new, chunk->map, old_size); 397 398 chunk->map_alloc = new_alloc; 399 chunk->map = new; 400 new = NULL; 401 402 out_unlock: 403 spin_unlock_irqrestore(&pcpu_lock, flags); 404 405 /* 406 * pcpu_mem_free() might end up calling vfree() which uses 407 * IRQ-unsafe lock and thus can't be called under pcpu_lock. 408 */ 409 pcpu_mem_free(old, old_size); 410 pcpu_mem_free(new, new_size); 411 412 return 0; 413 } 414 415 /** 416 * pcpu_split_block - split a map block 417 * @chunk: chunk of interest 418 * @i: index of map block to split 419 * @head: head size in bytes (can be 0) 420 * @tail: tail size in bytes (can be 0) 421 * 422 * Split the @i'th map block into two or three blocks. If @head is 423 * non-zero, @head bytes block is inserted before block @i moving it 424 * to @i+1 and reducing its size by @head bytes. 425 * 426 * If @tail is non-zero, the target block, which can be @i or @i+1 427 * depending on @head, is reduced by @tail bytes and @tail byte block 428 * is inserted after the target block. 429 * 430 * @chunk->map must have enough free slots to accomodate the split. 431 * 432 * CONTEXT: 433 * pcpu_lock. 434 */ 435 static void pcpu_split_block(struct pcpu_chunk *chunk, int i, 436 int head, int tail) 437 { 438 int nr_extra = !!head + !!tail; 439 440 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); 441 442 /* insert new subblocks */ 443 memmove(&chunk->map[i + nr_extra], &chunk->map[i], 444 sizeof(chunk->map[0]) * (chunk->map_used - i)); 445 chunk->map_used += nr_extra; 446 447 if (head) { 448 chunk->map[i + 1] = chunk->map[i] - head; 449 chunk->map[i++] = head; 450 } 451 if (tail) { 452 chunk->map[i++] -= tail; 453 chunk->map[i] = tail; 454 } 455 } 456 457 /** 458 * pcpu_alloc_area - allocate area from a pcpu_chunk 459 * @chunk: chunk of interest 460 * @size: wanted size in bytes 461 * @align: wanted align 462 * 463 * Try to allocate @size bytes area aligned at @align from @chunk. 464 * Note that this function only allocates the offset. It doesn't 465 * populate or map the area. 466 * 467 * @chunk->map must have at least two free slots. 468 * 469 * CONTEXT: 470 * pcpu_lock. 471 * 472 * RETURNS: 473 * Allocated offset in @chunk on success, -1 if no matching area is 474 * found. 475 */ 476 static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) 477 { 478 int oslot = pcpu_chunk_slot(chunk); 479 int max_contig = 0; 480 int i, off; 481 482 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { 483 bool is_last = i + 1 == chunk->map_used; 484 int head, tail; 485 486 /* extra for alignment requirement */ 487 head = ALIGN(off, align) - off; 488 BUG_ON(i == 0 && head != 0); 489 490 if (chunk->map[i] < 0) 491 continue; 492 if (chunk->map[i] < head + size) { 493 max_contig = max(chunk->map[i], max_contig); 494 continue; 495 } 496 497 /* 498 * If head is small or the previous block is free, 499 * merge'em. Note that 'small' is defined as smaller 500 * than sizeof(int), which is very small but isn't too 501 * uncommon for percpu allocations. 502 */ 503 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { 504 if (chunk->map[i - 1] > 0) 505 chunk->map[i - 1] += head; 506 else { 507 chunk->map[i - 1] -= head; 508 chunk->free_size -= head; 509 } 510 chunk->map[i] -= head; 511 off += head; 512 head = 0; 513 } 514 515 /* if tail is small, just keep it around */ 516 tail = chunk->map[i] - head - size; 517 if (tail < sizeof(int)) 518 tail = 0; 519 520 /* split if warranted */ 521 if (head || tail) { 522 pcpu_split_block(chunk, i, head, tail); 523 if (head) { 524 i++; 525 off += head; 526 max_contig = max(chunk->map[i - 1], max_contig); 527 } 528 if (tail) 529 max_contig = max(chunk->map[i + 1], max_contig); 530 } 531 532 /* update hint and mark allocated */ 533 if (is_last) 534 chunk->contig_hint = max_contig; /* fully scanned */ 535 else 536 chunk->contig_hint = max(chunk->contig_hint, 537 max_contig); 538 539 chunk->free_size -= chunk->map[i]; 540 chunk->map[i] = -chunk->map[i]; 541 542 pcpu_chunk_relocate(chunk, oslot); 543 return off; 544 } 545 546 chunk->contig_hint = max_contig; /* fully scanned */ 547 pcpu_chunk_relocate(chunk, oslot); 548 549 /* tell the upper layer that this chunk has no matching area */ 550 return -1; 551 } 552 553 /** 554 * pcpu_free_area - free area to a pcpu_chunk 555 * @chunk: chunk of interest 556 * @freeme: offset of area to free 557 * 558 * Free area starting from @freeme to @chunk. Note that this function 559 * only modifies the allocation map. It doesn't depopulate or unmap 560 * the area. 561 * 562 * CONTEXT: 563 * pcpu_lock. 564 */ 565 static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) 566 { 567 int oslot = pcpu_chunk_slot(chunk); 568 int i, off; 569 570 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) 571 if (off == freeme) 572 break; 573 BUG_ON(off != freeme); 574 BUG_ON(chunk->map[i] > 0); 575 576 chunk->map[i] = -chunk->map[i]; 577 chunk->free_size += chunk->map[i]; 578 579 /* merge with previous? */ 580 if (i > 0 && chunk->map[i - 1] >= 0) { 581 chunk->map[i - 1] += chunk->map[i]; 582 chunk->map_used--; 583 memmove(&chunk->map[i], &chunk->map[i + 1], 584 (chunk->map_used - i) * sizeof(chunk->map[0])); 585 i--; 586 } 587 /* merge with next? */ 588 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { 589 chunk->map[i] += chunk->map[i + 1]; 590 chunk->map_used--; 591 memmove(&chunk->map[i + 1], &chunk->map[i + 2], 592 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); 593 } 594 595 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); 596 pcpu_chunk_relocate(chunk, oslot); 597 } 598 599 static struct pcpu_chunk *pcpu_alloc_chunk(void) 600 { 601 struct pcpu_chunk *chunk; 602 603 chunk = pcpu_mem_alloc(pcpu_chunk_struct_size); 604 if (!chunk) 605 return NULL; 606 607 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); 608 if (!chunk->map) { 609 kfree(chunk); 610 return NULL; 611 } 612 613 chunk->map_alloc = PCPU_DFL_MAP_ALLOC; 614 chunk->map[chunk->map_used++] = pcpu_unit_size; 615 616 INIT_LIST_HEAD(&chunk->list); 617 chunk->free_size = pcpu_unit_size; 618 chunk->contig_hint = pcpu_unit_size; 619 620 return chunk; 621 } 622 623 static void pcpu_free_chunk(struct pcpu_chunk *chunk) 624 { 625 if (!chunk) 626 return; 627 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); 628 kfree(chunk); 629 } 630 631 /* 632 * Chunk management implementation. 633 * 634 * To allow different implementations, chunk alloc/free and 635 * [de]population are implemented in a separate file which is pulled 636 * into this file and compiled together. The following functions 637 * should be implemented. 638 * 639 * pcpu_populate_chunk - populate the specified range of a chunk 640 * pcpu_depopulate_chunk - depopulate the specified range of a chunk 641 * pcpu_create_chunk - create a new chunk 642 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop 643 * pcpu_addr_to_page - translate address to physical address 644 * pcpu_verify_alloc_info - check alloc_info is acceptable during init 645 */ 646 static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); 647 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); 648 static struct pcpu_chunk *pcpu_create_chunk(void); 649 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); 650 static struct page *pcpu_addr_to_page(void *addr); 651 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); 652 653 #ifdef CONFIG_NEED_PER_CPU_KM 654 #include "percpu-km.c" 655 #else 656 #include "percpu-vm.c" 657 #endif 658 659 /** 660 * pcpu_chunk_addr_search - determine chunk containing specified address 661 * @addr: address for which the chunk needs to be determined. 662 * 663 * RETURNS: 664 * The address of the found chunk. 665 */ 666 static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 667 { 668 /* is it in the first chunk? */ 669 if (pcpu_addr_in_first_chunk(addr)) { 670 /* is it in the reserved area? */ 671 if (pcpu_addr_in_reserved_chunk(addr)) 672 return pcpu_reserved_chunk; 673 return pcpu_first_chunk; 674 } 675 676 /* 677 * The address is relative to unit0 which might be unused and 678 * thus unmapped. Offset the address to the unit space of the 679 * current processor before looking it up in the vmalloc 680 * space. Note that any possible cpu id can be used here, so 681 * there's no need to worry about preemption or cpu hotplug. 682 */ 683 addr += pcpu_unit_offsets[raw_smp_processor_id()]; 684 return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); 685 } 686 687 /** 688 * pcpu_alloc - the percpu allocator 689 * @size: size of area to allocate in bytes 690 * @align: alignment of area (max PAGE_SIZE) 691 * @reserved: allocate from the reserved chunk if available 692 * 693 * Allocate percpu area of @size bytes aligned at @align. 694 * 695 * CONTEXT: 696 * Does GFP_KERNEL allocation. 697 * 698 * RETURNS: 699 * Percpu pointer to the allocated area on success, NULL on failure. 700 */ 701 static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved) 702 { 703 static int warn_limit = 10; 704 struct pcpu_chunk *chunk; 705 const char *err; 706 int slot, off, new_alloc; 707 unsigned long flags; 708 709 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { 710 WARN(true, "illegal size (%zu) or align (%zu) for " 711 "percpu allocation\n", size, align); 712 return NULL; 713 } 714 715 mutex_lock(&pcpu_alloc_mutex); 716 spin_lock_irqsave(&pcpu_lock, flags); 717 718 /* serve reserved allocations from the reserved chunk if available */ 719 if (reserved && pcpu_reserved_chunk) { 720 chunk = pcpu_reserved_chunk; 721 722 if (size > chunk->contig_hint) { 723 err = "alloc from reserved chunk failed"; 724 goto fail_unlock; 725 } 726 727 while ((new_alloc = pcpu_need_to_extend(chunk))) { 728 spin_unlock_irqrestore(&pcpu_lock, flags); 729 if (pcpu_extend_area_map(chunk, new_alloc) < 0) { 730 err = "failed to extend area map of reserved chunk"; 731 goto fail_unlock_mutex; 732 } 733 spin_lock_irqsave(&pcpu_lock, flags); 734 } 735 736 off = pcpu_alloc_area(chunk, size, align); 737 if (off >= 0) 738 goto area_found; 739 740 err = "alloc from reserved chunk failed"; 741 goto fail_unlock; 742 } 743 744 restart: 745 /* search through normal chunks */ 746 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 747 list_for_each_entry(chunk, &pcpu_slot[slot], list) { 748 if (size > chunk->contig_hint) 749 continue; 750 751 new_alloc = pcpu_need_to_extend(chunk); 752 if (new_alloc) { 753 spin_unlock_irqrestore(&pcpu_lock, flags); 754 if (pcpu_extend_area_map(chunk, 755 new_alloc) < 0) { 756 err = "failed to extend area map"; 757 goto fail_unlock_mutex; 758 } 759 spin_lock_irqsave(&pcpu_lock, flags); 760 /* 761 * pcpu_lock has been dropped, need to 762 * restart cpu_slot list walking. 763 */ 764 goto restart; 765 } 766 767 off = pcpu_alloc_area(chunk, size, align); 768 if (off >= 0) 769 goto area_found; 770 } 771 } 772 773 /* hmmm... no space left, create a new chunk */ 774 spin_unlock_irqrestore(&pcpu_lock, flags); 775 776 chunk = pcpu_create_chunk(); 777 if (!chunk) { 778 err = "failed to allocate new chunk"; 779 goto fail_unlock_mutex; 780 } 781 782 spin_lock_irqsave(&pcpu_lock, flags); 783 pcpu_chunk_relocate(chunk, -1); 784 goto restart; 785 786 area_found: 787 spin_unlock_irqrestore(&pcpu_lock, flags); 788 789 /* populate, map and clear the area */ 790 if (pcpu_populate_chunk(chunk, off, size)) { 791 spin_lock_irqsave(&pcpu_lock, flags); 792 pcpu_free_area(chunk, off); 793 err = "failed to populate"; 794 goto fail_unlock; 795 } 796 797 mutex_unlock(&pcpu_alloc_mutex); 798 799 /* return address relative to base address */ 800 return __addr_to_pcpu_ptr(chunk->base_addr + off); 801 802 fail_unlock: 803 spin_unlock_irqrestore(&pcpu_lock, flags); 804 fail_unlock_mutex: 805 mutex_unlock(&pcpu_alloc_mutex); 806 if (warn_limit) { 807 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, " 808 "%s\n", size, align, err); 809 dump_stack(); 810 if (!--warn_limit) 811 pr_info("PERCPU: limit reached, disable warning\n"); 812 } 813 return NULL; 814 } 815 816 /** 817 * __alloc_percpu - allocate dynamic percpu area 818 * @size: size of area to allocate in bytes 819 * @align: alignment of area (max PAGE_SIZE) 820 * 821 * Allocate percpu area of @size bytes aligned at @align. Might 822 * sleep. Might trigger writeouts. 823 * 824 * CONTEXT: 825 * Does GFP_KERNEL allocation. 826 * 827 * RETURNS: 828 * Percpu pointer to the allocated area on success, NULL on failure. 829 */ 830 void __percpu *__alloc_percpu(size_t size, size_t align) 831 { 832 return pcpu_alloc(size, align, false); 833 } 834 EXPORT_SYMBOL_GPL(__alloc_percpu); 835 836 /** 837 * __alloc_reserved_percpu - allocate reserved percpu area 838 * @size: size of area to allocate in bytes 839 * @align: alignment of area (max PAGE_SIZE) 840 * 841 * Allocate percpu area of @size bytes aligned at @align from reserved 842 * percpu area if arch has set it up; otherwise, allocation is served 843 * from the same dynamic area. Might sleep. Might trigger writeouts. 844 * 845 * CONTEXT: 846 * Does GFP_KERNEL allocation. 847 * 848 * RETURNS: 849 * Percpu pointer to the allocated area on success, NULL on failure. 850 */ 851 void __percpu *__alloc_reserved_percpu(size_t size, size_t align) 852 { 853 return pcpu_alloc(size, align, true); 854 } 855 856 /** 857 * pcpu_reclaim - reclaim fully free chunks, workqueue function 858 * @work: unused 859 * 860 * Reclaim all fully free chunks except for the first one. 861 * 862 * CONTEXT: 863 * workqueue context. 864 */ 865 static void pcpu_reclaim(struct work_struct *work) 866 { 867 LIST_HEAD(todo); 868 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; 869 struct pcpu_chunk *chunk, *next; 870 871 mutex_lock(&pcpu_alloc_mutex); 872 spin_lock_irq(&pcpu_lock); 873 874 list_for_each_entry_safe(chunk, next, head, list) { 875 WARN_ON(chunk->immutable); 876 877 /* spare the first one */ 878 if (chunk == list_first_entry(head, struct pcpu_chunk, list)) 879 continue; 880 881 list_move(&chunk->list, &todo); 882 } 883 884 spin_unlock_irq(&pcpu_lock); 885 886 list_for_each_entry_safe(chunk, next, &todo, list) { 887 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size); 888 pcpu_destroy_chunk(chunk); 889 } 890 891 mutex_unlock(&pcpu_alloc_mutex); 892 } 893 894 /** 895 * free_percpu - free percpu area 896 * @ptr: pointer to area to free 897 * 898 * Free percpu area @ptr. 899 * 900 * CONTEXT: 901 * Can be called from atomic context. 902 */ 903 void free_percpu(void __percpu *ptr) 904 { 905 void *addr; 906 struct pcpu_chunk *chunk; 907 unsigned long flags; 908 int off; 909 910 if (!ptr) 911 return; 912 913 addr = __pcpu_ptr_to_addr(ptr); 914 915 spin_lock_irqsave(&pcpu_lock, flags); 916 917 chunk = pcpu_chunk_addr_search(addr); 918 off = addr - chunk->base_addr; 919 920 pcpu_free_area(chunk, off); 921 922 /* if there are more than one fully free chunks, wake up grim reaper */ 923 if (chunk->free_size == pcpu_unit_size) { 924 struct pcpu_chunk *pos; 925 926 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 927 if (pos != chunk) { 928 schedule_work(&pcpu_reclaim_work); 929 break; 930 } 931 } 932 933 spin_unlock_irqrestore(&pcpu_lock, flags); 934 } 935 EXPORT_SYMBOL_GPL(free_percpu); 936 937 /** 938 * is_kernel_percpu_address - test whether address is from static percpu area 939 * @addr: address to test 940 * 941 * Test whether @addr belongs to in-kernel static percpu area. Module 942 * static percpu areas are not considered. For those, use 943 * is_module_percpu_address(). 944 * 945 * RETURNS: 946 * %true if @addr is from in-kernel static percpu area, %false otherwise. 947 */ 948 bool is_kernel_percpu_address(unsigned long addr) 949 { 950 const size_t static_size = __per_cpu_end - __per_cpu_start; 951 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 952 unsigned int cpu; 953 954 for_each_possible_cpu(cpu) { 955 void *start = per_cpu_ptr(base, cpu); 956 957 if ((void *)addr >= start && (void *)addr < start + static_size) 958 return true; 959 } 960 return false; 961 } 962 963 /** 964 * per_cpu_ptr_to_phys - convert translated percpu address to physical address 965 * @addr: the address to be converted to physical address 966 * 967 * Given @addr which is dereferenceable address obtained via one of 968 * percpu access macros, this function translates it into its physical 969 * address. The caller is responsible for ensuring @addr stays valid 970 * until this function finishes. 971 * 972 * RETURNS: 973 * The physical address for @addr. 974 */ 975 phys_addr_t per_cpu_ptr_to_phys(void *addr) 976 { 977 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 978 bool in_first_chunk = false; 979 unsigned long first_start, first_end; 980 unsigned int cpu; 981 982 /* 983 * The following test on first_start/end isn't strictly 984 * necessary but will speed up lookups of addresses which 985 * aren't in the first chunk. 986 */ 987 first_start = pcpu_chunk_addr(pcpu_first_chunk, pcpu_first_unit_cpu, 0); 988 first_end = pcpu_chunk_addr(pcpu_first_chunk, pcpu_last_unit_cpu, 989 pcpu_unit_pages); 990 if ((unsigned long)addr >= first_start && 991 (unsigned long)addr < first_end) { 992 for_each_possible_cpu(cpu) { 993 void *start = per_cpu_ptr(base, cpu); 994 995 if (addr >= start && addr < start + pcpu_unit_size) { 996 in_first_chunk = true; 997 break; 998 } 999 } 1000 } 1001 1002 if (in_first_chunk) { 1003 if ((unsigned long)addr < VMALLOC_START || 1004 (unsigned long)addr >= VMALLOC_END) 1005 return __pa(addr); 1006 else 1007 return page_to_phys(vmalloc_to_page(addr)); 1008 } else 1009 return page_to_phys(pcpu_addr_to_page(addr)); 1010 } 1011 1012 /** 1013 * pcpu_alloc_alloc_info - allocate percpu allocation info 1014 * @nr_groups: the number of groups 1015 * @nr_units: the number of units 1016 * 1017 * Allocate ai which is large enough for @nr_groups groups containing 1018 * @nr_units units. The returned ai's groups[0].cpu_map points to the 1019 * cpu_map array which is long enough for @nr_units and filled with 1020 * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1021 * pointer of other groups. 1022 * 1023 * RETURNS: 1024 * Pointer to the allocated pcpu_alloc_info on success, NULL on 1025 * failure. 1026 */ 1027 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 1028 int nr_units) 1029 { 1030 struct pcpu_alloc_info *ai; 1031 size_t base_size, ai_size; 1032 void *ptr; 1033 int unit; 1034 1035 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 1036 __alignof__(ai->groups[0].cpu_map[0])); 1037 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 1038 1039 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size)); 1040 if (!ptr) 1041 return NULL; 1042 ai = ptr; 1043 ptr += base_size; 1044 1045 ai->groups[0].cpu_map = ptr; 1046 1047 for (unit = 0; unit < nr_units; unit++) 1048 ai->groups[0].cpu_map[unit] = NR_CPUS; 1049 1050 ai->nr_groups = nr_groups; 1051 ai->__ai_size = PFN_ALIGN(ai_size); 1052 1053 return ai; 1054 } 1055 1056 /** 1057 * pcpu_free_alloc_info - free percpu allocation info 1058 * @ai: pcpu_alloc_info to free 1059 * 1060 * Free @ai which was allocated by pcpu_alloc_alloc_info(). 1061 */ 1062 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 1063 { 1064 free_bootmem(__pa(ai), ai->__ai_size); 1065 } 1066 1067 /** 1068 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 1069 * @reserved_size: the size of reserved percpu area in bytes 1070 * @dyn_size: minimum free size for dynamic allocation in bytes 1071 * @atom_size: allocation atom size 1072 * @cpu_distance_fn: callback to determine distance between cpus, optional 1073 * 1074 * This function determines grouping of units, their mappings to cpus 1075 * and other parameters considering needed percpu size, allocation 1076 * atom size and distances between CPUs. 1077 * 1078 * Groups are always mutliples of atom size and CPUs which are of 1079 * LOCAL_DISTANCE both ways are grouped together and share space for 1080 * units in the same group. The returned configuration is guaranteed 1081 * to have CPUs on different nodes on different groups and >=75% usage 1082 * of allocated virtual address space. 1083 * 1084 * RETURNS: 1085 * On success, pointer to the new allocation_info is returned. On 1086 * failure, ERR_PTR value is returned. 1087 */ 1088 static struct pcpu_alloc_info * __init pcpu_build_alloc_info( 1089 size_t reserved_size, size_t dyn_size, 1090 size_t atom_size, 1091 pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 1092 { 1093 static int group_map[NR_CPUS] __initdata; 1094 static int group_cnt[NR_CPUS] __initdata; 1095 const size_t static_size = __per_cpu_end - __per_cpu_start; 1096 int nr_groups = 1, nr_units = 0; 1097 size_t size_sum, min_unit_size, alloc_size; 1098 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 1099 int last_allocs, group, unit; 1100 unsigned int cpu, tcpu; 1101 struct pcpu_alloc_info *ai; 1102 unsigned int *cpu_map; 1103 1104 /* this function may be called multiple times */ 1105 memset(group_map, 0, sizeof(group_map)); 1106 memset(group_cnt, 0, sizeof(group_cnt)); 1107 1108 /* calculate size_sum and ensure dyn_size is enough for early alloc */ 1109 size_sum = PFN_ALIGN(static_size + reserved_size + 1110 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); 1111 dyn_size = size_sum - static_size - reserved_size; 1112 1113 /* 1114 * Determine min_unit_size, alloc_size and max_upa such that 1115 * alloc_size is multiple of atom_size and is the smallest 1116 * which can accomodate 4k aligned segments which are equal to 1117 * or larger than min_unit_size. 1118 */ 1119 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 1120 1121 alloc_size = roundup(min_unit_size, atom_size); 1122 upa = alloc_size / min_unit_size; 1123 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1124 upa--; 1125 max_upa = upa; 1126 1127 /* group cpus according to their proximity */ 1128 for_each_possible_cpu(cpu) { 1129 group = 0; 1130 next_group: 1131 for_each_possible_cpu(tcpu) { 1132 if (cpu == tcpu) 1133 break; 1134 if (group_map[tcpu] == group && cpu_distance_fn && 1135 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 1136 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 1137 group++; 1138 nr_groups = max(nr_groups, group + 1); 1139 goto next_group; 1140 } 1141 } 1142 group_map[cpu] = group; 1143 group_cnt[group]++; 1144 } 1145 1146 /* 1147 * Expand unit size until address space usage goes over 75% 1148 * and then as much as possible without using more address 1149 * space. 1150 */ 1151 last_allocs = INT_MAX; 1152 for (upa = max_upa; upa; upa--) { 1153 int allocs = 0, wasted = 0; 1154 1155 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1156 continue; 1157 1158 for (group = 0; group < nr_groups; group++) { 1159 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 1160 allocs += this_allocs; 1161 wasted += this_allocs * upa - group_cnt[group]; 1162 } 1163 1164 /* 1165 * Don't accept if wastage is over 25%. The 1166 * greater-than comparison ensures upa==1 always 1167 * passes the following check. 1168 */ 1169 if (wasted > num_possible_cpus() / 3) 1170 continue; 1171 1172 /* and then don't consume more memory */ 1173 if (allocs > last_allocs) 1174 break; 1175 last_allocs = allocs; 1176 best_upa = upa; 1177 } 1178 upa = best_upa; 1179 1180 /* allocate and fill alloc_info */ 1181 for (group = 0; group < nr_groups; group++) 1182 nr_units += roundup(group_cnt[group], upa); 1183 1184 ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 1185 if (!ai) 1186 return ERR_PTR(-ENOMEM); 1187 cpu_map = ai->groups[0].cpu_map; 1188 1189 for (group = 0; group < nr_groups; group++) { 1190 ai->groups[group].cpu_map = cpu_map; 1191 cpu_map += roundup(group_cnt[group], upa); 1192 } 1193 1194 ai->static_size = static_size; 1195 ai->reserved_size = reserved_size; 1196 ai->dyn_size = dyn_size; 1197 ai->unit_size = alloc_size / upa; 1198 ai->atom_size = atom_size; 1199 ai->alloc_size = alloc_size; 1200 1201 for (group = 0, unit = 0; group_cnt[group]; group++) { 1202 struct pcpu_group_info *gi = &ai->groups[group]; 1203 1204 /* 1205 * Initialize base_offset as if all groups are located 1206 * back-to-back. The caller should update this to 1207 * reflect actual allocation. 1208 */ 1209 gi->base_offset = unit * ai->unit_size; 1210 1211 for_each_possible_cpu(cpu) 1212 if (group_map[cpu] == group) 1213 gi->cpu_map[gi->nr_units++] = cpu; 1214 gi->nr_units = roundup(gi->nr_units, upa); 1215 unit += gi->nr_units; 1216 } 1217 BUG_ON(unit != nr_units); 1218 1219 return ai; 1220 } 1221 1222 /** 1223 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 1224 * @lvl: loglevel 1225 * @ai: allocation info to dump 1226 * 1227 * Print out information about @ai using loglevel @lvl. 1228 */ 1229 static void pcpu_dump_alloc_info(const char *lvl, 1230 const struct pcpu_alloc_info *ai) 1231 { 1232 int group_width = 1, cpu_width = 1, width; 1233 char empty_str[] = "--------"; 1234 int alloc = 0, alloc_end = 0; 1235 int group, v; 1236 int upa, apl; /* units per alloc, allocs per line */ 1237 1238 v = ai->nr_groups; 1239 while (v /= 10) 1240 group_width++; 1241 1242 v = num_possible_cpus(); 1243 while (v /= 10) 1244 cpu_width++; 1245 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 1246 1247 upa = ai->alloc_size / ai->unit_size; 1248 width = upa * (cpu_width + 1) + group_width + 3; 1249 apl = rounddown_pow_of_two(max(60 / width, 1)); 1250 1251 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 1252 lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 1253 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 1254 1255 for (group = 0; group < ai->nr_groups; group++) { 1256 const struct pcpu_group_info *gi = &ai->groups[group]; 1257 int unit = 0, unit_end = 0; 1258 1259 BUG_ON(gi->nr_units % upa); 1260 for (alloc_end += gi->nr_units / upa; 1261 alloc < alloc_end; alloc++) { 1262 if (!(alloc % apl)) { 1263 printk("\n"); 1264 printk("%spcpu-alloc: ", lvl); 1265 } 1266 printk("[%0*d] ", group_width, group); 1267 1268 for (unit_end += upa; unit < unit_end; unit++) 1269 if (gi->cpu_map[unit] != NR_CPUS) 1270 printk("%0*d ", cpu_width, 1271 gi->cpu_map[unit]); 1272 else 1273 printk("%s ", empty_str); 1274 } 1275 } 1276 printk("\n"); 1277 } 1278 1279 /** 1280 * pcpu_setup_first_chunk - initialize the first percpu chunk 1281 * @ai: pcpu_alloc_info describing how to percpu area is shaped 1282 * @base_addr: mapped address 1283 * 1284 * Initialize the first percpu chunk which contains the kernel static 1285 * perpcu area. This function is to be called from arch percpu area 1286 * setup path. 1287 * 1288 * @ai contains all information necessary to initialize the first 1289 * chunk and prime the dynamic percpu allocator. 1290 * 1291 * @ai->static_size is the size of static percpu area. 1292 * 1293 * @ai->reserved_size, if non-zero, specifies the amount of bytes to 1294 * reserve after the static area in the first chunk. This reserves 1295 * the first chunk such that it's available only through reserved 1296 * percpu allocation. This is primarily used to serve module percpu 1297 * static areas on architectures where the addressing model has 1298 * limited offset range for symbol relocations to guarantee module 1299 * percpu symbols fall inside the relocatable range. 1300 * 1301 * @ai->dyn_size determines the number of bytes available for dynamic 1302 * allocation in the first chunk. The area between @ai->static_size + 1303 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 1304 * 1305 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 1306 * and equal to or larger than @ai->static_size + @ai->reserved_size + 1307 * @ai->dyn_size. 1308 * 1309 * @ai->atom_size is the allocation atom size and used as alignment 1310 * for vm areas. 1311 * 1312 * @ai->alloc_size is the allocation size and always multiple of 1313 * @ai->atom_size. This is larger than @ai->atom_size if 1314 * @ai->unit_size is larger than @ai->atom_size. 1315 * 1316 * @ai->nr_groups and @ai->groups describe virtual memory layout of 1317 * percpu areas. Units which should be colocated are put into the 1318 * same group. Dynamic VM areas will be allocated according to these 1319 * groupings. If @ai->nr_groups is zero, a single group containing 1320 * all units is assumed. 1321 * 1322 * The caller should have mapped the first chunk at @base_addr and 1323 * copied static data to each unit. 1324 * 1325 * If the first chunk ends up with both reserved and dynamic areas, it 1326 * is served by two chunks - one to serve the core static and reserved 1327 * areas and the other for the dynamic area. They share the same vm 1328 * and page map but uses different area allocation map to stay away 1329 * from each other. The latter chunk is circulated in the chunk slots 1330 * and available for dynamic allocation like any other chunks. 1331 * 1332 * RETURNS: 1333 * 0 on success, -errno on failure. 1334 */ 1335 int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 1336 void *base_addr) 1337 { 1338 static char cpus_buf[4096] __initdata; 1339 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; 1340 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; 1341 size_t dyn_size = ai->dyn_size; 1342 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; 1343 struct pcpu_chunk *schunk, *dchunk = NULL; 1344 unsigned long *group_offsets; 1345 size_t *group_sizes; 1346 unsigned long *unit_off; 1347 unsigned int cpu; 1348 int *unit_map; 1349 int group, unit, i; 1350 1351 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask); 1352 1353 #define PCPU_SETUP_BUG_ON(cond) do { \ 1354 if (unlikely(cond)) { \ 1355 pr_emerg("PERCPU: failed to initialize, %s", #cond); \ 1356 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \ 1357 pcpu_dump_alloc_info(KERN_EMERG, ai); \ 1358 BUG(); \ 1359 } \ 1360 } while (0) 1361 1362 /* sanity checks */ 1363 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 1364 PCPU_SETUP_BUG_ON(!ai->static_size); 1365 PCPU_SETUP_BUG_ON(!base_addr); 1366 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 1367 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK); 1368 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 1369 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); 1370 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); 1371 1372 /* process group information and build config tables accordingly */ 1373 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0])); 1374 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0])); 1375 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0])); 1376 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0])); 1377 1378 for (cpu = 0; cpu < nr_cpu_ids; cpu++) 1379 unit_map[cpu] = UINT_MAX; 1380 pcpu_first_unit_cpu = NR_CPUS; 1381 1382 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 1383 const struct pcpu_group_info *gi = &ai->groups[group]; 1384 1385 group_offsets[group] = gi->base_offset; 1386 group_sizes[group] = gi->nr_units * ai->unit_size; 1387 1388 for (i = 0; i < gi->nr_units; i++) { 1389 cpu = gi->cpu_map[i]; 1390 if (cpu == NR_CPUS) 1391 continue; 1392 1393 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids); 1394 PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 1395 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 1396 1397 unit_map[cpu] = unit + i; 1398 unit_off[cpu] = gi->base_offset + i * ai->unit_size; 1399 1400 if (pcpu_first_unit_cpu == NR_CPUS) 1401 pcpu_first_unit_cpu = cpu; 1402 } 1403 } 1404 pcpu_last_unit_cpu = cpu; 1405 pcpu_nr_units = unit; 1406 1407 for_each_possible_cpu(cpu) 1408 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 1409 1410 /* we're done parsing the input, undefine BUG macro and dump config */ 1411 #undef PCPU_SETUP_BUG_ON 1412 pcpu_dump_alloc_info(KERN_INFO, ai); 1413 1414 pcpu_nr_groups = ai->nr_groups; 1415 pcpu_group_offsets = group_offsets; 1416 pcpu_group_sizes = group_sizes; 1417 pcpu_unit_map = unit_map; 1418 pcpu_unit_offsets = unit_off; 1419 1420 /* determine basic parameters */ 1421 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 1422 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 1423 pcpu_atom_size = ai->atom_size; 1424 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 1425 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 1426 1427 /* 1428 * Allocate chunk slots. The additional last slot is for 1429 * empty chunks. 1430 */ 1431 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 1432 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); 1433 for (i = 0; i < pcpu_nr_slots; i++) 1434 INIT_LIST_HEAD(&pcpu_slot[i]); 1435 1436 /* 1437 * Initialize static chunk. If reserved_size is zero, the 1438 * static chunk covers static area + dynamic allocation area 1439 * in the first chunk. If reserved_size is not zero, it 1440 * covers static area + reserved area (mostly used for module 1441 * static percpu allocation). 1442 */ 1443 schunk = alloc_bootmem(pcpu_chunk_struct_size); 1444 INIT_LIST_HEAD(&schunk->list); 1445 schunk->base_addr = base_addr; 1446 schunk->map = smap; 1447 schunk->map_alloc = ARRAY_SIZE(smap); 1448 schunk->immutable = true; 1449 bitmap_fill(schunk->populated, pcpu_unit_pages); 1450 1451 if (ai->reserved_size) { 1452 schunk->free_size = ai->reserved_size; 1453 pcpu_reserved_chunk = schunk; 1454 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; 1455 } else { 1456 schunk->free_size = dyn_size; 1457 dyn_size = 0; /* dynamic area covered */ 1458 } 1459 schunk->contig_hint = schunk->free_size; 1460 1461 schunk->map[schunk->map_used++] = -ai->static_size; 1462 if (schunk->free_size) 1463 schunk->map[schunk->map_used++] = schunk->free_size; 1464 1465 /* init dynamic chunk if necessary */ 1466 if (dyn_size) { 1467 dchunk = alloc_bootmem(pcpu_chunk_struct_size); 1468 INIT_LIST_HEAD(&dchunk->list); 1469 dchunk->base_addr = base_addr; 1470 dchunk->map = dmap; 1471 dchunk->map_alloc = ARRAY_SIZE(dmap); 1472 dchunk->immutable = true; 1473 bitmap_fill(dchunk->populated, pcpu_unit_pages); 1474 1475 dchunk->contig_hint = dchunk->free_size = dyn_size; 1476 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; 1477 dchunk->map[dchunk->map_used++] = dchunk->free_size; 1478 } 1479 1480 /* link the first chunk in */ 1481 pcpu_first_chunk = dchunk ?: schunk; 1482 pcpu_chunk_relocate(pcpu_first_chunk, -1); 1483 1484 /* we're done */ 1485 pcpu_base_addr = base_addr; 1486 return 0; 1487 } 1488 1489 const char *pcpu_fc_names[PCPU_FC_NR] __initdata = { 1490 [PCPU_FC_AUTO] = "auto", 1491 [PCPU_FC_EMBED] = "embed", 1492 [PCPU_FC_PAGE] = "page", 1493 }; 1494 1495 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 1496 1497 static int __init percpu_alloc_setup(char *str) 1498 { 1499 if (0) 1500 /* nada */; 1501 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 1502 else if (!strcmp(str, "embed")) 1503 pcpu_chosen_fc = PCPU_FC_EMBED; 1504 #endif 1505 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1506 else if (!strcmp(str, "page")) 1507 pcpu_chosen_fc = PCPU_FC_PAGE; 1508 #endif 1509 else 1510 pr_warning("PERCPU: unknown allocator %s specified\n", str); 1511 1512 return 0; 1513 } 1514 early_param("percpu_alloc", percpu_alloc_setup); 1515 1516 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 1517 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 1518 /** 1519 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 1520 * @reserved_size: the size of reserved percpu area in bytes 1521 * @dyn_size: minimum free size for dynamic allocation in bytes 1522 * @atom_size: allocation atom size 1523 * @cpu_distance_fn: callback to determine distance between cpus, optional 1524 * @alloc_fn: function to allocate percpu page 1525 * @free_fn: funtion to free percpu page 1526 * 1527 * This is a helper to ease setting up embedded first percpu chunk and 1528 * can be called where pcpu_setup_first_chunk() is expected. 1529 * 1530 * If this function is used to setup the first chunk, it is allocated 1531 * by calling @alloc_fn and used as-is without being mapped into 1532 * vmalloc area. Allocations are always whole multiples of @atom_size 1533 * aligned to @atom_size. 1534 * 1535 * This enables the first chunk to piggy back on the linear physical 1536 * mapping which often uses larger page size. Please note that this 1537 * can result in very sparse cpu->unit mapping on NUMA machines thus 1538 * requiring large vmalloc address space. Don't use this allocator if 1539 * vmalloc space is not orders of magnitude larger than distances 1540 * between node memory addresses (ie. 32bit NUMA machines). 1541 * 1542 * @dyn_size specifies the minimum dynamic area size. 1543 * 1544 * If the needed size is smaller than the minimum or specified unit 1545 * size, the leftover is returned using @free_fn. 1546 * 1547 * RETURNS: 1548 * 0 on success, -errno on failure. 1549 */ 1550 int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, 1551 size_t atom_size, 1552 pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 1553 pcpu_fc_alloc_fn_t alloc_fn, 1554 pcpu_fc_free_fn_t free_fn) 1555 { 1556 void *base = (void *)ULONG_MAX; 1557 void **areas = NULL; 1558 struct pcpu_alloc_info *ai; 1559 size_t size_sum, areas_size, max_distance; 1560 int group, i, rc; 1561 1562 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 1563 cpu_distance_fn); 1564 if (IS_ERR(ai)) 1565 return PTR_ERR(ai); 1566 1567 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 1568 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 1569 1570 areas = alloc_bootmem_nopanic(areas_size); 1571 if (!areas) { 1572 rc = -ENOMEM; 1573 goto out_free; 1574 } 1575 1576 /* allocate, copy and determine base address */ 1577 for (group = 0; group < ai->nr_groups; group++) { 1578 struct pcpu_group_info *gi = &ai->groups[group]; 1579 unsigned int cpu = NR_CPUS; 1580 void *ptr; 1581 1582 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 1583 cpu = gi->cpu_map[i]; 1584 BUG_ON(cpu == NR_CPUS); 1585 1586 /* allocate space for the whole group */ 1587 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 1588 if (!ptr) { 1589 rc = -ENOMEM; 1590 goto out_free_areas; 1591 } 1592 areas[group] = ptr; 1593 1594 base = min(ptr, base); 1595 1596 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 1597 if (gi->cpu_map[i] == NR_CPUS) { 1598 /* unused unit, free whole */ 1599 free_fn(ptr, ai->unit_size); 1600 continue; 1601 } 1602 /* copy and return the unused part */ 1603 memcpy(ptr, __per_cpu_load, ai->static_size); 1604 free_fn(ptr + size_sum, ai->unit_size - size_sum); 1605 } 1606 } 1607 1608 /* base address is now known, determine group base offsets */ 1609 max_distance = 0; 1610 for (group = 0; group < ai->nr_groups; group++) { 1611 ai->groups[group].base_offset = areas[group] - base; 1612 max_distance = max_t(size_t, max_distance, 1613 ai->groups[group].base_offset); 1614 } 1615 max_distance += ai->unit_size; 1616 1617 /* warn if maximum distance is further than 75% of vmalloc space */ 1618 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) { 1619 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc " 1620 "space 0x%lx\n", 1621 max_distance, VMALLOC_END - VMALLOC_START); 1622 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1623 /* and fail if we have fallback */ 1624 rc = -EINVAL; 1625 goto out_free; 1626 #endif 1627 } 1628 1629 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 1630 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 1631 ai->dyn_size, ai->unit_size); 1632 1633 rc = pcpu_setup_first_chunk(ai, base); 1634 goto out_free; 1635 1636 out_free_areas: 1637 for (group = 0; group < ai->nr_groups; group++) 1638 free_fn(areas[group], 1639 ai->groups[group].nr_units * ai->unit_size); 1640 out_free: 1641 pcpu_free_alloc_info(ai); 1642 if (areas) 1643 free_bootmem(__pa(areas), areas_size); 1644 return rc; 1645 } 1646 #endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK || 1647 !CONFIG_HAVE_SETUP_PER_CPU_AREA */ 1648 1649 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1650 /** 1651 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 1652 * @reserved_size: the size of reserved percpu area in bytes 1653 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 1654 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE 1655 * @populate_pte_fn: function to populate pte 1656 * 1657 * This is a helper to ease setting up page-remapped first percpu 1658 * chunk and can be called where pcpu_setup_first_chunk() is expected. 1659 * 1660 * This is the basic allocator. Static percpu area is allocated 1661 * page-by-page into vmalloc area. 1662 * 1663 * RETURNS: 1664 * 0 on success, -errno on failure. 1665 */ 1666 int __init pcpu_page_first_chunk(size_t reserved_size, 1667 pcpu_fc_alloc_fn_t alloc_fn, 1668 pcpu_fc_free_fn_t free_fn, 1669 pcpu_fc_populate_pte_fn_t populate_pte_fn) 1670 { 1671 static struct vm_struct vm; 1672 struct pcpu_alloc_info *ai; 1673 char psize_str[16]; 1674 int unit_pages; 1675 size_t pages_size; 1676 struct page **pages; 1677 int unit, i, j, rc; 1678 1679 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 1680 1681 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); 1682 if (IS_ERR(ai)) 1683 return PTR_ERR(ai); 1684 BUG_ON(ai->nr_groups != 1); 1685 BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); 1686 1687 unit_pages = ai->unit_size >> PAGE_SHIFT; 1688 1689 /* unaligned allocations can't be freed, round up to page size */ 1690 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 1691 sizeof(pages[0])); 1692 pages = alloc_bootmem(pages_size); 1693 1694 /* allocate pages */ 1695 j = 0; 1696 for (unit = 0; unit < num_possible_cpus(); unit++) 1697 for (i = 0; i < unit_pages; i++) { 1698 unsigned int cpu = ai->groups[0].cpu_map[unit]; 1699 void *ptr; 1700 1701 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 1702 if (!ptr) { 1703 pr_warning("PERCPU: failed to allocate %s page " 1704 "for cpu%u\n", psize_str, cpu); 1705 goto enomem; 1706 } 1707 pages[j++] = virt_to_page(ptr); 1708 } 1709 1710 /* allocate vm area, map the pages and copy static data */ 1711 vm.flags = VM_ALLOC; 1712 vm.size = num_possible_cpus() * ai->unit_size; 1713 vm_area_register_early(&vm, PAGE_SIZE); 1714 1715 for (unit = 0; unit < num_possible_cpus(); unit++) { 1716 unsigned long unit_addr = 1717 (unsigned long)vm.addr + unit * ai->unit_size; 1718 1719 for (i = 0; i < unit_pages; i++) 1720 populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 1721 1722 /* pte already populated, the following shouldn't fail */ 1723 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 1724 unit_pages); 1725 if (rc < 0) 1726 panic("failed to map percpu area, err=%d\n", rc); 1727 1728 /* 1729 * FIXME: Archs with virtual cache should flush local 1730 * cache for the linear mapping here - something 1731 * equivalent to flush_cache_vmap() on the local cpu. 1732 * flush_cache_vmap() can't be used as most supporting 1733 * data structures are not set up yet. 1734 */ 1735 1736 /* copy static data */ 1737 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 1738 } 1739 1740 /* we're ready, commit */ 1741 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n", 1742 unit_pages, psize_str, vm.addr, ai->static_size, 1743 ai->reserved_size, ai->dyn_size); 1744 1745 rc = pcpu_setup_first_chunk(ai, vm.addr); 1746 goto out_free_ar; 1747 1748 enomem: 1749 while (--j >= 0) 1750 free_fn(page_address(pages[j]), PAGE_SIZE); 1751 rc = -ENOMEM; 1752 out_free_ar: 1753 free_bootmem(__pa(pages), pages_size); 1754 pcpu_free_alloc_info(ai); 1755 return rc; 1756 } 1757 #endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */ 1758 1759 /* 1760 * Generic percpu area setup. 1761 * 1762 * The embedding helper is used because its behavior closely resembles 1763 * the original non-dynamic generic percpu area setup. This is 1764 * important because many archs have addressing restrictions and might 1765 * fail if the percpu area is located far away from the previous 1766 * location. As an added bonus, in non-NUMA cases, embedding is 1767 * generally a good idea TLB-wise because percpu area can piggy back 1768 * on the physical linear memory mapping which uses large page 1769 * mappings on applicable archs. 1770 */ 1771 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 1772 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 1773 EXPORT_SYMBOL(__per_cpu_offset); 1774 1775 static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 1776 size_t align) 1777 { 1778 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS)); 1779 } 1780 1781 static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 1782 { 1783 free_bootmem(__pa(ptr), size); 1784 } 1785 1786 void __init setup_per_cpu_areas(void) 1787 { 1788 unsigned long delta; 1789 unsigned int cpu; 1790 int rc; 1791 1792 /* 1793 * Always reserve area for module percpu variables. That's 1794 * what the legacy allocator did. 1795 */ 1796 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 1797 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 1798 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 1799 if (rc < 0) 1800 panic("Failed to initialized percpu areas."); 1801 1802 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 1803 for_each_possible_cpu(cpu) 1804 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 1805 } 1806 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 1807 1808 /* 1809 * First and reserved chunks are initialized with temporary allocation 1810 * map in initdata so that they can be used before slab is online. 1811 * This function is called after slab is brought up and replaces those 1812 * with properly allocated maps. 1813 */ 1814 void __init percpu_init_late(void) 1815 { 1816 struct pcpu_chunk *target_chunks[] = 1817 { pcpu_first_chunk, pcpu_reserved_chunk, NULL }; 1818 struct pcpu_chunk *chunk; 1819 unsigned long flags; 1820 int i; 1821 1822 for (i = 0; (chunk = target_chunks[i]); i++) { 1823 int *map; 1824 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]); 1825 1826 BUILD_BUG_ON(size > PAGE_SIZE); 1827 1828 map = pcpu_mem_alloc(size); 1829 BUG_ON(!map); 1830 1831 spin_lock_irqsave(&pcpu_lock, flags); 1832 memcpy(map, chunk->map, size); 1833 chunk->map = map; 1834 spin_unlock_irqrestore(&pcpu_lock, flags); 1835 } 1836 } 1837