~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/mm/percpu.c

Version: ~ [ linux-5.1-rc5 ] ~ [ linux-5.0.7 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.34 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.111 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.168 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.178 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.138 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.65 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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 addresses */
114 static unsigned int pcpu_low_unit_cpu __read_mostly;
115 static unsigned int pcpu_high_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 int __pcpu_size_to_slot(int size)
180 {
181         int highbit = fls(size);        /* size is in bytes */
182         return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
183 }
184 
185 static int pcpu_size_to_slot(int size)
186 {
187         if (size == pcpu_unit_size)
188                 return pcpu_nr_slots - 1;
189         return __pcpu_size_to_slot(size);
190 }
191 
192 static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
193 {
194         if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
195                 return 0;
196 
197         return pcpu_size_to_slot(chunk->free_size);
198 }
199 
200 static int pcpu_page_idx(unsigned int cpu, int page_idx)
201 {
202         return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
203 }
204 
205 static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
206                                      unsigned int cpu, int page_idx)
207 {
208         return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
209                 (page_idx << PAGE_SHIFT);
210 }
211 
212 static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
213                                     unsigned int cpu, int page_idx)
214 {
215         /* must not be used on pre-mapped chunk */
216         WARN_ON(chunk->immutable);
217 
218         return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
219 }
220 
221 /* set the pointer to a chunk in a page struct */
222 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
223 {
224         page->index = (unsigned long)pcpu;
225 }
226 
227 /* obtain pointer to a chunk from a page struct */
228 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
229 {
230         return (struct pcpu_chunk *)page->index;
231 }
232 
233 static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
234 {
235         *rs = find_next_zero_bit(chunk->populated, end, *rs);
236         *re = find_next_bit(chunk->populated, end, *rs + 1);
237 }
238 
239 static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
240 {
241         *rs = find_next_bit(chunk->populated, end, *rs);
242         *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
243 }
244 
245 /*
246  * (Un)populated page region iterators.  Iterate over (un)populated
247  * page regions betwen @start and @end in @chunk.  @rs and @re should
248  * be integer variables and will be set to start and end page index of
249  * the current region.
250  */
251 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end)               \
252         for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
253              (rs) < (re);                                                   \
254              (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
255 
256 #define pcpu_for_each_pop_region(chunk, rs, re, start, end)                 \
257         for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end));   \
258              (rs) < (re);                                                   \
259              (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
260 
261 /**
262  * pcpu_mem_alloc - allocate memory
263  * @size: bytes to allocate
264  *
265  * Allocate @size bytes.  If @size is smaller than PAGE_SIZE,
266  * kzalloc() is used; otherwise, vmalloc() is used.  The returned
267  * memory is always zeroed.
268  *
269  * CONTEXT:
270  * Does GFP_KERNEL allocation.
271  *
272  * RETURNS:
273  * Pointer to the allocated area on success, NULL on failure.
274  */
275 static void *pcpu_mem_alloc(size_t size)
276 {
277         if (size <= PAGE_SIZE)
278                 return kzalloc(size, GFP_KERNEL);
279         else {
280                 void *ptr = vmalloc(size);
281                 if (ptr)
282                         memset(ptr, 0, size);
283                 return ptr;
284         }
285 }
286 
287 /**
288  * pcpu_mem_free - free memory
289  * @ptr: memory to free
290  * @size: size of the area
291  *
292  * Free @ptr.  @ptr should have been allocated using pcpu_mem_alloc().
293  */
294 static void pcpu_mem_free(void *ptr, size_t size)
295 {
296         if (size <= PAGE_SIZE)
297                 kfree(ptr);
298         else
299                 vfree(ptr);
300 }
301 
302 /**
303  * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
304  * @chunk: chunk of interest
305  * @oslot: the previous slot it was on
306  *
307  * This function is called after an allocation or free changed @chunk.
308  * New slot according to the changed state is determined and @chunk is
309  * moved to the slot.  Note that the reserved chunk is never put on
310  * chunk slots.
311  *
312  * CONTEXT:
313  * pcpu_lock.
314  */
315 static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
316 {
317         int nslot = pcpu_chunk_slot(chunk);
318 
319         if (chunk != pcpu_reserved_chunk && oslot != nslot) {
320                 if (oslot < nslot)
321                         list_move(&chunk->list, &pcpu_slot[nslot]);
322                 else
323                         list_move_tail(&chunk->list, &pcpu_slot[nslot]);
324         }
325 }
326 
327 /**
328  * pcpu_chunk_addr_search - determine chunk containing specified address
329  * @addr: address for which the chunk needs to be determined.
330  *
331  * RETURNS:
332  * The address of the found chunk.
333  */
334 static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
335 {
336         void *first_start = pcpu_first_chunk->base_addr;
337 
338         /* is it in the first chunk? */
339         if (addr >= first_start && addr < first_start + pcpu_unit_size) {
340                 /* is it in the reserved area? */
341                 if (addr < first_start + pcpu_reserved_chunk_limit)
342                         return pcpu_reserved_chunk;
343                 return pcpu_first_chunk;
344         }
345 
346         /*
347          * The address is relative to unit0 which might be unused and
348          * thus unmapped.  Offset the address to the unit space of the
349          * current processor before looking it up in the vmalloc
350          * space.  Note that any possible cpu id can be used here, so
351          * there's no need to worry about preemption or cpu hotplug.
352          */
353         addr += pcpu_unit_offsets[raw_smp_processor_id()];
354         return pcpu_get_page_chunk(vmalloc_to_page(addr));
355 }
356 
357 /**
358  * pcpu_need_to_extend - determine whether chunk area map needs to be extended
359  * @chunk: chunk of interest
360  *
361  * Determine whether area map of @chunk needs to be extended to
362  * accomodate a new allocation.
363  *
364  * CONTEXT:
365  * pcpu_lock.
366  *
367  * RETURNS:
368  * New target map allocation length if extension is necessary, 0
369  * otherwise.
370  */
371 static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
372 {
373         int new_alloc;
374 
375         if (chunk->map_alloc >= chunk->map_used + 2)
376                 return 0;
377 
378         new_alloc = PCPU_DFL_MAP_ALLOC;
379         while (new_alloc < chunk->map_used + 2)
380                 new_alloc *= 2;
381 
382         return new_alloc;
383 }
384 
385 /**
386  * pcpu_extend_area_map - extend area map of a chunk
387  * @chunk: chunk of interest
388  * @new_alloc: new target allocation length of the area map
389  *
390  * Extend area map of @chunk to have @new_alloc entries.
391  *
392  * CONTEXT:
393  * Does GFP_KERNEL allocation.  Grabs and releases pcpu_lock.
394  *
395  * RETURNS:
396  * 0 on success, -errno on failure.
397  */
398 static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
399 {
400         int *old = NULL, *new = NULL;
401         size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
402         unsigned long flags;
403 
404         new = pcpu_mem_alloc(new_size);
405         if (!new)
406                 return -ENOMEM;
407 
408         /* acquire pcpu_lock and switch to new area map */
409         spin_lock_irqsave(&pcpu_lock, flags);
410 
411         if (new_alloc <= chunk->map_alloc)
412                 goto out_unlock;
413 
414         old_size = chunk->map_alloc * sizeof(chunk->map[0]);
415         memcpy(new, chunk->map, old_size);
416 
417         /*
418          * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
419          * one of the first chunks and still using static map.
420          */
421         if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
422                 old = chunk->map;
423 
424         chunk->map_alloc = new_alloc;
425         chunk->map = new;
426         new = NULL;
427 
428 out_unlock:
429         spin_unlock_irqrestore(&pcpu_lock, flags);
430 
431         /*
432          * pcpu_mem_free() might end up calling vfree() which uses
433          * IRQ-unsafe lock and thus can't be called under pcpu_lock.
434          */
435         pcpu_mem_free(old, old_size);
436         pcpu_mem_free(new, new_size);
437 
438         return 0;
439 }
440 
441 /**
442  * pcpu_split_block - split a map block
443  * @chunk: chunk of interest
444  * @i: index of map block to split
445  * @head: head size in bytes (can be 0)
446  * @tail: tail size in bytes (can be 0)
447  *
448  * Split the @i'th map block into two or three blocks.  If @head is
449  * non-zero, @head bytes block is inserted before block @i moving it
450  * to @i+1 and reducing its size by @head bytes.
451  *
452  * If @tail is non-zero, the target block, which can be @i or @i+1
453  * depending on @head, is reduced by @tail bytes and @tail byte block
454  * is inserted after the target block.
455  *
456  * @chunk->map must have enough free slots to accomodate the split.
457  *
458  * CONTEXT:
459  * pcpu_lock.
460  */
461 static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
462                              int head, int tail)
463 {
464         int nr_extra = !!head + !!tail;
465 
466         BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
467 
468         /* insert new subblocks */
469         memmove(&chunk->map[i + nr_extra], &chunk->map[i],
470                 sizeof(chunk->map[0]) * (chunk->map_used - i));
471         chunk->map_used += nr_extra;
472 
473         if (head) {
474                 chunk->map[i + 1] = chunk->map[i] - head;
475                 chunk->map[i++] = head;
476         }
477         if (tail) {
478                 chunk->map[i++] -= tail;
479                 chunk->map[i] = tail;
480         }
481 }
482 
483 /**
484  * pcpu_alloc_area - allocate area from a pcpu_chunk
485  * @chunk: chunk of interest
486  * @size: wanted size in bytes
487  * @align: wanted align
488  *
489  * Try to allocate @size bytes area aligned at @align from @chunk.
490  * Note that this function only allocates the offset.  It doesn't
491  * populate or map the area.
492  *
493  * @chunk->map must have at least two free slots.
494  *
495  * CONTEXT:
496  * pcpu_lock.
497  *
498  * RETURNS:
499  * Allocated offset in @chunk on success, -1 if no matching area is
500  * found.
501  */
502 static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
503 {
504         int oslot = pcpu_chunk_slot(chunk);
505         int max_contig = 0;
506         int i, off;
507 
508         for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
509                 bool is_last = i + 1 == chunk->map_used;
510                 int head, tail;
511 
512                 /* extra for alignment requirement */
513                 head = ALIGN(off, align) - off;
514                 BUG_ON(i == 0 && head != 0);
515 
516                 if (chunk->map[i] < 0)
517                         continue;
518                 if (chunk->map[i] < head + size) {
519                         max_contig = max(chunk->map[i], max_contig);
520                         continue;
521                 }
522 
523                 /*
524                  * If head is small or the previous block is free,
525                  * merge'em.  Note that 'small' is defined as smaller
526                  * than sizeof(int), which is very small but isn't too
527                  * uncommon for percpu allocations.
528                  */
529                 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
530                         if (chunk->map[i - 1] > 0)
531                                 chunk->map[i - 1] += head;
532                         else {
533                                 chunk->map[i - 1] -= head;
534                                 chunk->free_size -= head;
535                         }
536                         chunk->map[i] -= head;
537                         off += head;
538                         head = 0;
539                 }
540 
541                 /* if tail is small, just keep it around */
542                 tail = chunk->map[i] - head - size;
543                 if (tail < sizeof(int))
544                         tail = 0;
545 
546                 /* split if warranted */
547                 if (head || tail) {
548                         pcpu_split_block(chunk, i, head, tail);
549                         if (head) {
550                                 i++;
551                                 off += head;
552                                 max_contig = max(chunk->map[i - 1], max_contig);
553                         }
554                         if (tail)
555                                 max_contig = max(chunk->map[i + 1], max_contig);
556                 }
557 
558                 /* update hint and mark allocated */
559                 if (is_last)
560                         chunk->contig_hint = max_contig; /* fully scanned */
561                 else
562                         chunk->contig_hint = max(chunk->contig_hint,
563                                                  max_contig);
564 
565                 chunk->free_size -= chunk->map[i];
566                 chunk->map[i] = -chunk->map[i];
567 
568                 pcpu_chunk_relocate(chunk, oslot);
569                 return off;
570         }
571 
572         chunk->contig_hint = max_contig;        /* fully scanned */
573         pcpu_chunk_relocate(chunk, oslot);
574 
575         /* tell the upper layer that this chunk has no matching area */
576         return -1;
577 }
578 
579 /**
580  * pcpu_free_area - free area to a pcpu_chunk
581  * @chunk: chunk of interest
582  * @freeme: offset of area to free
583  *
584  * Free area starting from @freeme to @chunk.  Note that this function
585  * only modifies the allocation map.  It doesn't depopulate or unmap
586  * the area.
587  *
588  * CONTEXT:
589  * pcpu_lock.
590  */
591 static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
592 {
593         int oslot = pcpu_chunk_slot(chunk);
594         int i, off;
595 
596         for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
597                 if (off == freeme)
598                         break;
599         BUG_ON(off != freeme);
600         BUG_ON(chunk->map[i] > 0);
601 
602         chunk->map[i] = -chunk->map[i];
603         chunk->free_size += chunk->map[i];
604 
605         /* merge with previous? */
606         if (i > 0 && chunk->map[i - 1] >= 0) {
607                 chunk->map[i - 1] += chunk->map[i];
608                 chunk->map_used--;
609                 memmove(&chunk->map[i], &chunk->map[i + 1],
610                         (chunk->map_used - i) * sizeof(chunk->map[0]));
611                 i--;
612         }
613         /* merge with next? */
614         if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
615                 chunk->map[i] += chunk->map[i + 1];
616                 chunk->map_used--;
617                 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
618                         (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
619         }
620 
621         chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
622         pcpu_chunk_relocate(chunk, oslot);
623 }
624 
625 /**
626  * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
627  * @chunk: chunk of interest
628  * @bitmapp: output parameter for bitmap
629  * @may_alloc: may allocate the array
630  *
631  * Returns pointer to array of pointers to struct page and bitmap,
632  * both of which can be indexed with pcpu_page_idx().  The returned
633  * array is cleared to zero and *@bitmapp is copied from
634  * @chunk->populated.  Note that there is only one array and bitmap
635  * and access exclusion is the caller's responsibility.
636  *
637  * CONTEXT:
638  * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
639  * Otherwise, don't care.
640  *
641  * RETURNS:
642  * Pointer to temp pages array on success, NULL on failure.
643  */
644 static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
645                                                unsigned long **bitmapp,
646                                                bool may_alloc)
647 {
648         static struct page **pages;
649         static unsigned long *bitmap;
650         size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
651         size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
652                              sizeof(unsigned long);
653 
654         if (!pages || !bitmap) {
655                 if (may_alloc && !pages)
656                         pages = pcpu_mem_alloc(pages_size);
657                 if (may_alloc && !bitmap)
658                         bitmap = pcpu_mem_alloc(bitmap_size);
659                 if (!pages || !bitmap)
660                         return NULL;
661         }
662 
663         memset(pages, 0, pages_size);
664         bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
665 
666         *bitmapp = bitmap;
667         return pages;
668 }
669 
670 /**
671  * pcpu_free_pages - free pages which were allocated for @chunk
672  * @chunk: chunk pages were allocated for
673  * @pages: array of pages to be freed, indexed by pcpu_page_idx()
674  * @populated: populated bitmap
675  * @page_start: page index of the first page to be freed
676  * @page_end: page index of the last page to be freed + 1
677  *
678  * Free pages [@page_start and @page_end) in @pages for all units.
679  * The pages were allocated for @chunk.
680  */
681 static void pcpu_free_pages(struct pcpu_chunk *chunk,
682                             struct page **pages, unsigned long *populated,
683                             int page_start, int page_end)
684 {
685         unsigned int cpu;
686         int i;
687 
688         for_each_possible_cpu(cpu) {
689                 for (i = page_start; i < page_end; i++) {
690                         struct page *page = pages[pcpu_page_idx(cpu, i)];
691 
692                         if (page)
693                                 __free_page(page);
694                 }
695         }
696 }
697 
698 /**
699  * pcpu_alloc_pages - allocates pages for @chunk
700  * @chunk: target chunk
701  * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
702  * @populated: populated bitmap
703  * @page_start: page index of the first page to be allocated
704  * @page_end: page index of the last page to be allocated + 1
705  *
706  * Allocate pages [@page_start,@page_end) into @pages for all units.
707  * The allocation is for @chunk.  Percpu core doesn't care about the
708  * content of @pages and will pass it verbatim to pcpu_map_pages().
709  */
710 static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
711                             struct page **pages, unsigned long *populated,
712                             int page_start, int page_end)
713 {
714         const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
715         unsigned int cpu;
716         int i;
717 
718         for_each_possible_cpu(cpu) {
719                 for (i = page_start; i < page_end; i++) {
720                         struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
721 
722                         *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
723                         if (!*pagep) {
724                                 pcpu_free_pages(chunk, pages, populated,
725                                                 page_start, page_end);
726                                 return -ENOMEM;
727                         }
728                 }
729         }
730         return 0;
731 }
732 
733 /**
734  * pcpu_pre_unmap_flush - flush cache prior to unmapping
735  * @chunk: chunk the regions to be flushed belongs to
736  * @page_start: page index of the first page to be flushed
737  * @page_end: page index of the last page to be flushed + 1
738  *
739  * Pages in [@page_start,@page_end) of @chunk are about to be
740  * unmapped.  Flush cache.  As each flushing trial can be very
741  * expensive, issue flush on the whole region at once rather than
742  * doing it for each cpu.  This could be an overkill but is more
743  * scalable.
744  */
745 static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
746                                  int page_start, int page_end)
747 {
748         flush_cache_vunmap(
749                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
750                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
751 }
752 
753 static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
754 {
755         unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
756 }
757 
758 /**
759  * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
760  * @chunk: chunk of interest
761  * @pages: pages array which can be used to pass information to free
762  * @populated: populated bitmap
763  * @page_start: page index of the first page to unmap
764  * @page_end: page index of the last page to unmap + 1
765  *
766  * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
767  * Corresponding elements in @pages were cleared by the caller and can
768  * be used to carry information to pcpu_free_pages() which will be
769  * called after all unmaps are finished.  The caller should call
770  * proper pre/post flush functions.
771  */
772 static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
773                              struct page **pages, unsigned long *populated,
774                              int page_start, int page_end)
775 {
776         unsigned int cpu;
777         int i;
778 
779         for_each_possible_cpu(cpu) {
780                 for (i = page_start; i < page_end; i++) {
781                         struct page *page;
782 
783                         page = pcpu_chunk_page(chunk, cpu, i);
784                         WARN_ON(!page);
785                         pages[pcpu_page_idx(cpu, i)] = page;
786                 }
787                 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
788                                    page_end - page_start);
789         }
790 
791         for (i = page_start; i < page_end; i++)
792                 __clear_bit(i, populated);
793 }
794 
795 /**
796  * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
797  * @chunk: pcpu_chunk the regions to be flushed belong to
798  * @page_start: page index of the first page to be flushed
799  * @page_end: page index of the last page to be flushed + 1
800  *
801  * Pages [@page_start,@page_end) of @chunk have been unmapped.  Flush
802  * TLB for the regions.  This can be skipped if the area is to be
803  * returned to vmalloc as vmalloc will handle TLB flushing lazily.
804  *
805  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
806  * for the whole region.
807  */
808 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
809                                       int page_start, int page_end)
810 {
811         flush_tlb_kernel_range(
812                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
813                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
814 }
815 
816 static int __pcpu_map_pages(unsigned long addr, struct page **pages,
817                             int nr_pages)
818 {
819         return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
820                                         PAGE_KERNEL, pages);
821 }
822 
823 /**
824  * pcpu_map_pages - map pages into a pcpu_chunk
825  * @chunk: chunk of interest
826  * @pages: pages array containing pages to be mapped
827  * @populated: populated bitmap
828  * @page_start: page index of the first page to map
829  * @page_end: page index of the last page to map + 1
830  *
831  * For each cpu, map pages [@page_start,@page_end) into @chunk.  The
832  * caller is responsible for calling pcpu_post_map_flush() after all
833  * mappings are complete.
834  *
835  * This function is responsible for setting corresponding bits in
836  * @chunk->populated bitmap and whatever is necessary for reverse
837  * lookup (addr -> chunk).
838  */
839 static int pcpu_map_pages(struct pcpu_chunk *chunk,
840                           struct page **pages, unsigned long *populated,
841                           int page_start, int page_end)
842 {
843         unsigned int cpu, tcpu;
844         int i, err;
845 
846         for_each_possible_cpu(cpu) {
847                 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
848                                        &pages[pcpu_page_idx(cpu, page_start)],
849                                        page_end - page_start);
850                 if (err < 0)
851                         goto err;
852         }
853 
854         /* mapping successful, link chunk and mark populated */
855         for (i = page_start; i < page_end; i++) {
856                 for_each_possible_cpu(cpu)
857                         pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
858                                             chunk);
859                 __set_bit(i, populated);
860         }
861 
862         return 0;
863 
864 err:
865         for_each_possible_cpu(tcpu) {
866                 if (tcpu == cpu)
867                         break;
868                 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
869                                    page_end - page_start);
870         }
871         return err;
872 }
873 
874 /**
875  * pcpu_post_map_flush - flush cache after mapping
876  * @chunk: pcpu_chunk the regions to be flushed belong to
877  * @page_start: page index of the first page to be flushed
878  * @page_end: page index of the last page to be flushed + 1
879  *
880  * Pages [@page_start,@page_end) of @chunk have been mapped.  Flush
881  * cache.
882  *
883  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
884  * for the whole region.
885  */
886 static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
887                                 int page_start, int page_end)
888 {
889         flush_cache_vmap(
890                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
891                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
892 }
893 
894 /**
895  * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
896  * @chunk: chunk to depopulate
897  * @off: offset to the area to depopulate
898  * @size: size of the area to depopulate in bytes
899  * @flush: whether to flush cache and tlb or not
900  *
901  * For each cpu, depopulate and unmap pages [@page_start,@page_end)
902  * from @chunk.  If @flush is true, vcache is flushed before unmapping
903  * and tlb after.
904  *
905  * CONTEXT:
906  * pcpu_alloc_mutex.
907  */
908 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
909 {
910         int page_start = PFN_DOWN(off);
911         int page_end = PFN_UP(off + size);
912         struct page **pages;
913         unsigned long *populated;
914         int rs, re;
915 
916         /* quick path, check whether it's empty already */
917         pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
918                 if (rs == page_start && re == page_end)
919                         return;
920                 break;
921         }
922 
923         /* immutable chunks can't be depopulated */
924         WARN_ON(chunk->immutable);
925 
926         /*
927          * If control reaches here, there must have been at least one
928          * successful population attempt so the temp pages array must
929          * be available now.
930          */
931         pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
932         BUG_ON(!pages);
933 
934         /* unmap and free */
935         pcpu_pre_unmap_flush(chunk, page_start, page_end);
936 
937         pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
938                 pcpu_unmap_pages(chunk, pages, populated, rs, re);
939 
940         /* no need to flush tlb, vmalloc will handle it lazily */
941 
942         pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
943                 pcpu_free_pages(chunk, pages, populated, rs, re);
944 
945         /* commit new bitmap */
946         bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
947 }
948 
949 /**
950  * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
951  * @chunk: chunk of interest
952  * @off: offset to the area to populate
953  * @size: size of the area to populate in bytes
954  *
955  * For each cpu, populate and map pages [@page_start,@page_end) into
956  * @chunk.  The area is cleared on return.
957  *
958  * CONTEXT:
959  * pcpu_alloc_mutex, does GFP_KERNEL allocation.
960  */
961 static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
962 {
963         int page_start = PFN_DOWN(off);
964         int page_end = PFN_UP(off + size);
965         int free_end = page_start, unmap_end = page_start;
966         struct page **pages;
967         unsigned long *populated;
968         unsigned int cpu;
969         int rs, re, rc;
970 
971         /* quick path, check whether all pages are already there */
972         pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) {
973                 if (rs == page_start && re == page_end)
974                         goto clear;
975                 break;
976         }
977 
978         /* need to allocate and map pages, this chunk can't be immutable */
979         WARN_ON(chunk->immutable);
980 
981         pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
982         if (!pages)
983                 return -ENOMEM;
984 
985         /* alloc and map */
986         pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
987                 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
988                 if (rc)
989                         goto err_free;
990                 free_end = re;
991         }
992 
993         pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
994                 rc = pcpu_map_pages(chunk, pages, populated, rs, re);
995                 if (rc)
996                         goto err_unmap;
997                 unmap_end = re;
998         }
999         pcpu_post_map_flush(chunk, page_start, page_end);
1000 
1001         /* commit new bitmap */
1002         bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
1003 clear:
1004         for_each_possible_cpu(cpu)
1005                 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1006         return 0;
1007 
1008 err_unmap:
1009         pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
1010         pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
1011                 pcpu_unmap_pages(chunk, pages, populated, rs, re);
1012         pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
1013 err_free:
1014         pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
1015                 pcpu_free_pages(chunk, pages, populated, rs, re);
1016         return rc;
1017 }
1018 
1019 static void free_pcpu_chunk(struct pcpu_chunk *chunk)
1020 {
1021         if (!chunk)
1022                 return;
1023         if (chunk->vms)
1024                 pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups);
1025         pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
1026         kfree(chunk);
1027 }
1028 
1029 static struct pcpu_chunk *alloc_pcpu_chunk(void)
1030 {
1031         struct pcpu_chunk *chunk;
1032 
1033         chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
1034         if (!chunk)
1035                 return NULL;
1036 
1037         chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
1038         chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
1039         chunk->map[chunk->map_used++] = pcpu_unit_size;
1040 
1041         chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
1042                                        pcpu_nr_groups, pcpu_atom_size,
1043                                        GFP_KERNEL);
1044         if (!chunk->vms) {
1045                 free_pcpu_chunk(chunk);
1046                 return NULL;
1047         }
1048 
1049         INIT_LIST_HEAD(&chunk->list);
1050         chunk->free_size = pcpu_unit_size;
1051         chunk->contig_hint = pcpu_unit_size;
1052         chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0];
1053 
1054         return chunk;
1055 }
1056 
1057 /**
1058  * pcpu_alloc - the percpu allocator
1059  * @size: size of area to allocate in bytes
1060  * @align: alignment of area (max PAGE_SIZE)
1061  * @reserved: allocate from the reserved chunk if available
1062  *
1063  * Allocate percpu area of @size bytes aligned at @align.
1064  *
1065  * CONTEXT:
1066  * Does GFP_KERNEL allocation.
1067  *
1068  * RETURNS:
1069  * Percpu pointer to the allocated area on success, NULL on failure.
1070  */
1071 static void *pcpu_alloc(size_t size, size_t align, bool reserved)
1072 {
1073         static int warn_limit = 10;
1074         struct pcpu_chunk *chunk;
1075         const char *err;
1076         int slot, off, new_alloc;
1077         unsigned long flags;
1078 
1079         if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
1080                 WARN(true, "illegal size (%zu) or align (%zu) for "
1081                      "percpu allocation\n", size, align);
1082                 return NULL;
1083         }
1084 
1085         mutex_lock(&pcpu_alloc_mutex);
1086         spin_lock_irqsave(&pcpu_lock, flags);
1087 
1088         /* serve reserved allocations from the reserved chunk if available */
1089         if (reserved && pcpu_reserved_chunk) {
1090                 chunk = pcpu_reserved_chunk;
1091 
1092                 if (size > chunk->contig_hint) {
1093                         err = "alloc from reserved chunk failed";
1094                         goto fail_unlock;
1095                 }
1096 
1097                 while ((new_alloc = pcpu_need_to_extend(chunk))) {
1098                         spin_unlock_irqrestore(&pcpu_lock, flags);
1099                         if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
1100                                 err = "failed to extend area map of reserved chunk";
1101                                 goto fail_unlock_mutex;
1102                         }
1103                         spin_lock_irqsave(&pcpu_lock, flags);
1104                 }
1105 
1106                 off = pcpu_alloc_area(chunk, size, align);
1107                 if (off >= 0)
1108                         goto area_found;
1109 
1110                 err = "alloc from reserved chunk failed";
1111                 goto fail_unlock;
1112         }
1113 
1114 restart:
1115         /* search through normal chunks */
1116         for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1117                 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1118                         if (size > chunk->contig_hint)
1119                                 continue;
1120 
1121                         new_alloc = pcpu_need_to_extend(chunk);
1122                         if (new_alloc) {
1123                                 spin_unlock_irqrestore(&pcpu_lock, flags);
1124                                 if (pcpu_extend_area_map(chunk,
1125                                                          new_alloc) < 0) {
1126                                         err = "failed to extend area map";
1127                                         goto fail_unlock_mutex;
1128                                 }
1129                                 spin_lock_irqsave(&pcpu_lock, flags);
1130                                 /*
1131                                  * pcpu_lock has been dropped, need to
1132                                  * restart cpu_slot list walking.
1133                                  */
1134                                 goto restart;
1135                         }
1136 
1137                         off = pcpu_alloc_area(chunk, size, align);
1138                         if (off >= 0)
1139                                 goto area_found;
1140                 }
1141         }
1142 
1143         /* hmmm... no space left, create a new chunk */
1144         spin_unlock_irqrestore(&pcpu_lock, flags);
1145 
1146         chunk = alloc_pcpu_chunk();
1147         if (!chunk) {
1148                 err = "failed to allocate new chunk";
1149                 goto fail_unlock_mutex;
1150         }
1151 
1152         spin_lock_irqsave(&pcpu_lock, flags);
1153         pcpu_chunk_relocate(chunk, -1);
1154         goto restart;
1155 
1156 area_found:
1157         spin_unlock_irqrestore(&pcpu_lock, flags);
1158 
1159         /* populate, map and clear the area */
1160         if (pcpu_populate_chunk(chunk, off, size)) {
1161                 spin_lock_irqsave(&pcpu_lock, flags);
1162                 pcpu_free_area(chunk, off);
1163                 err = "failed to populate";
1164                 goto fail_unlock;
1165         }
1166 
1167         mutex_unlock(&pcpu_alloc_mutex);
1168 
1169         /* return address relative to base address */
1170         return __addr_to_pcpu_ptr(chunk->base_addr + off);
1171 
1172 fail_unlock:
1173         spin_unlock_irqrestore(&pcpu_lock, flags);
1174 fail_unlock_mutex:
1175         mutex_unlock(&pcpu_alloc_mutex);
1176         if (warn_limit) {
1177                 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
1178                            "%s\n", size, align, err);
1179                 dump_stack();
1180                 if (!--warn_limit)
1181                         pr_info("PERCPU: limit reached, disable warning\n");
1182         }
1183         return NULL;
1184 }
1185 
1186 /**
1187  * __alloc_percpu - allocate dynamic percpu area
1188  * @size: size of area to allocate in bytes
1189  * @align: alignment of area (max PAGE_SIZE)
1190  *
1191  * Allocate percpu area of @size bytes aligned at @align.  Might
1192  * sleep.  Might trigger writeouts.
1193  *
1194  * CONTEXT:
1195  * Does GFP_KERNEL allocation.
1196  *
1197  * RETURNS:
1198  * Percpu pointer to the allocated area on success, NULL on failure.
1199  */
1200 void *__alloc_percpu(size_t size, size_t align)
1201 {
1202         return pcpu_alloc(size, align, false);
1203 }
1204 EXPORT_SYMBOL_GPL(__alloc_percpu);
1205 
1206 /**
1207  * __alloc_reserved_percpu - allocate reserved percpu area
1208  * @size: size of area to allocate in bytes
1209  * @align: alignment of area (max PAGE_SIZE)
1210  *
1211  * Allocate percpu area of @size bytes aligned at @align from reserved
1212  * percpu area if arch has set it up; otherwise, allocation is served
1213  * from the same dynamic area.  Might sleep.  Might trigger writeouts.
1214  *
1215  * CONTEXT:
1216  * Does GFP_KERNEL allocation.
1217  *
1218  * RETURNS:
1219  * Percpu pointer to the allocated area on success, NULL on failure.
1220  */
1221 void *__alloc_reserved_percpu(size_t size, size_t align)
1222 {
1223         return pcpu_alloc(size, align, true);
1224 }
1225 
1226 /**
1227  * pcpu_reclaim - reclaim fully free chunks, workqueue function
1228  * @work: unused
1229  *
1230  * Reclaim all fully free chunks except for the first one.
1231  *
1232  * CONTEXT:
1233  * workqueue context.
1234  */
1235 static void pcpu_reclaim(struct work_struct *work)
1236 {
1237         LIST_HEAD(todo);
1238         struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
1239         struct pcpu_chunk *chunk, *next;
1240 
1241         mutex_lock(&pcpu_alloc_mutex);
1242         spin_lock_irq(&pcpu_lock);
1243 
1244         list_for_each_entry_safe(chunk, next, head, list) {
1245                 WARN_ON(chunk->immutable);
1246 
1247                 /* spare the first one */
1248                 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
1249                         continue;
1250 
1251                 list_move(&chunk->list, &todo);
1252         }
1253 
1254         spin_unlock_irq(&pcpu_lock);
1255 
1256         list_for_each_entry_safe(chunk, next, &todo, list) {
1257                 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
1258                 free_pcpu_chunk(chunk);
1259         }
1260 
1261         mutex_unlock(&pcpu_alloc_mutex);
1262 }
1263 
1264 /**
1265  * free_percpu - free percpu area
1266  * @ptr: pointer to area to free
1267  *
1268  * Free percpu area @ptr.
1269  *
1270  * CONTEXT:
1271  * Can be called from atomic context.
1272  */
1273 void free_percpu(void *ptr)
1274 {
1275         void *addr = __pcpu_ptr_to_addr(ptr);
1276         struct pcpu_chunk *chunk;
1277         unsigned long flags;
1278         int off;
1279 
1280         if (!ptr)
1281                 return;
1282 
1283         spin_lock_irqsave(&pcpu_lock, flags);
1284 
1285         chunk = pcpu_chunk_addr_search(addr);
1286         off = addr - chunk->base_addr;
1287 
1288         pcpu_free_area(chunk, off);
1289 
1290         /* if there are more than one fully free chunks, wake up grim reaper */
1291         if (chunk->free_size == pcpu_unit_size) {
1292                 struct pcpu_chunk *pos;
1293 
1294                 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1295                         if (pos != chunk) {
1296                                 schedule_work(&pcpu_reclaim_work);
1297                                 break;
1298                         }
1299         }
1300 
1301         spin_unlock_irqrestore(&pcpu_lock, flags);
1302 }
1303 EXPORT_SYMBOL_GPL(free_percpu);
1304 
1305 static inline size_t pcpu_calc_fc_sizes(size_t static_size,
1306                                         size_t reserved_size,
1307                                         ssize_t *dyn_sizep)
1308 {
1309         size_t size_sum;
1310 
1311         size_sum = PFN_ALIGN(static_size + reserved_size +
1312                              (*dyn_sizep >= 0 ? *dyn_sizep : 0));
1313         if (*dyn_sizep != 0)
1314                 *dyn_sizep = size_sum - static_size - reserved_size;
1315 
1316         return size_sum;
1317 }
1318 
1319 /**
1320  * pcpu_alloc_alloc_info - allocate percpu allocation info
1321  * @nr_groups: the number of groups
1322  * @nr_units: the number of units
1323  *
1324  * Allocate ai which is large enough for @nr_groups groups containing
1325  * @nr_units units.  The returned ai's groups[0].cpu_map points to the
1326  * cpu_map array which is long enough for @nr_units and filled with
1327  * NR_CPUS.  It's the caller's responsibility to initialize cpu_map
1328  * pointer of other groups.
1329  *
1330  * RETURNS:
1331  * Pointer to the allocated pcpu_alloc_info on success, NULL on
1332  * failure.
1333  */
1334 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1335                                                       int nr_units)
1336 {
1337         struct pcpu_alloc_info *ai;
1338         size_t base_size, ai_size;
1339         void *ptr;
1340         int unit;
1341 
1342         base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1343                           __alignof__(ai->groups[0].cpu_map[0]));
1344         ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1345 
1346         ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1347         if (!ptr)
1348                 return NULL;
1349         ai = ptr;
1350         ptr += base_size;
1351 
1352         ai->groups[0].cpu_map = ptr;
1353 
1354         for (unit = 0; unit < nr_units; unit++)
1355                 ai->groups[0].cpu_map[unit] = NR_CPUS;
1356 
1357         ai->nr_groups = nr_groups;
1358         ai->__ai_size = PFN_ALIGN(ai_size);
1359 
1360         return ai;
1361 }
1362 
1363 /**
1364  * pcpu_free_alloc_info - free percpu allocation info
1365  * @ai: pcpu_alloc_info to free
1366  *
1367  * Free @ai which was allocated by pcpu_alloc_alloc_info().
1368  */
1369 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1370 {
1371         free_bootmem(__pa(ai), ai->__ai_size);
1372 }
1373 
1374 /**
1375  * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1376  * @reserved_size: the size of reserved percpu area in bytes
1377  * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1378  * @atom_size: allocation atom size
1379  * @cpu_distance_fn: callback to determine distance between cpus, optional
1380  *
1381  * This function determines grouping of units, their mappings to cpus
1382  * and other parameters considering needed percpu size, allocation
1383  * atom size and distances between CPUs.
1384  *
1385  * Groups are always mutliples of atom size and CPUs which are of
1386  * LOCAL_DISTANCE both ways are grouped together and share space for
1387  * units in the same group.  The returned configuration is guaranteed
1388  * to have CPUs on different nodes on different groups and >=75% usage
1389  * of allocated virtual address space.
1390  *
1391  * RETURNS:
1392  * On success, pointer to the new allocation_info is returned.  On
1393  * failure, ERR_PTR value is returned.
1394  */
1395 struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1396                                 size_t reserved_size, ssize_t dyn_size,
1397                                 size_t atom_size,
1398                                 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1399 {
1400         static int group_map[NR_CPUS] __initdata;
1401         static int group_cnt[NR_CPUS] __initdata;
1402         const size_t static_size = __per_cpu_end - __per_cpu_start;
1403         int group_cnt_max = 0, nr_groups = 1, nr_units = 0;
1404         size_t size_sum, min_unit_size, alloc_size;
1405         int upa, max_upa, uninitialized_var(best_upa);  /* units_per_alloc */
1406         int last_allocs, group, unit;
1407         unsigned int cpu, tcpu;
1408         struct pcpu_alloc_info *ai;
1409         unsigned int *cpu_map;
1410 
1411         /* this function may be called multiple times */
1412         memset(group_map, 0, sizeof(group_map));
1413         memset(group_cnt, 0, sizeof(group_map));
1414 
1415         /*
1416          * Determine min_unit_size, alloc_size and max_upa such that
1417          * alloc_size is multiple of atom_size and is the smallest
1418          * which can accomodate 4k aligned segments which are equal to
1419          * or larger than min_unit_size.
1420          */
1421         size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size);
1422         min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1423 
1424         alloc_size = roundup(min_unit_size, atom_size);
1425         upa = alloc_size / min_unit_size;
1426         while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1427                 upa--;
1428         max_upa = upa;
1429 
1430         /* group cpus according to their proximity */
1431         for_each_possible_cpu(cpu) {
1432                 group = 0;
1433         next_group:
1434                 for_each_possible_cpu(tcpu) {
1435                         if (cpu == tcpu)
1436                                 break;
1437                         if (group_map[tcpu] == group && cpu_distance_fn &&
1438                             (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1439                              cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1440                                 group++;
1441                                 nr_groups = max(nr_groups, group + 1);
1442                                 goto next_group;
1443                         }
1444                 }
1445                 group_map[cpu] = group;
1446                 group_cnt[group]++;
1447                 group_cnt_max = max(group_cnt_max, group_cnt[group]);
1448         }
1449 
1450         /*
1451          * Expand unit size until address space usage goes over 75%
1452          * and then as much as possible without using more address
1453          * space.
1454          */
1455         last_allocs = INT_MAX;
1456         for (upa = max_upa; upa; upa--) {
1457                 int allocs = 0, wasted = 0;
1458 
1459                 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1460                         continue;
1461 
1462                 for (group = 0; group < nr_groups; group++) {
1463                         int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1464                         allocs += this_allocs;
1465                         wasted += this_allocs * upa - group_cnt[group];
1466                 }
1467 
1468                 /*
1469                  * Don't accept if wastage is over 25%.  The
1470                  * greater-than comparison ensures upa==1 always
1471                  * passes the following check.
1472                  */
1473                 if (wasted > num_possible_cpus() / 3)
1474                         continue;
1475 
1476                 /* and then don't consume more memory */
1477                 if (allocs > last_allocs)
1478                         break;
1479                 last_allocs = allocs;
1480                 best_upa = upa;
1481         }
1482         upa = best_upa;
1483 
1484         /* allocate and fill alloc_info */
1485         for (group = 0; group < nr_groups; group++)
1486                 nr_units += roundup(group_cnt[group], upa);
1487 
1488         ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1489         if (!ai)
1490                 return ERR_PTR(-ENOMEM);
1491         cpu_map = ai->groups[0].cpu_map;
1492 
1493         for (group = 0; group < nr_groups; group++) {
1494                 ai->groups[group].cpu_map = cpu_map;
1495                 cpu_map += roundup(group_cnt[group], upa);
1496         }
1497 
1498         ai->static_size = static_size;
1499         ai->reserved_size = reserved_size;
1500         ai->dyn_size = dyn_size;
1501         ai->unit_size = alloc_size / upa;
1502         ai->atom_size = atom_size;
1503         ai->alloc_size = alloc_size;
1504 
1505         for (group = 0, unit = 0; group_cnt[group]; group++) {
1506                 struct pcpu_group_info *gi = &ai->groups[group];
1507 
1508                 /*
1509                  * Initialize base_offset as if all groups are located
1510                  * back-to-back.  The caller should update this to
1511                  * reflect actual allocation.
1512                  */
1513                 gi->base_offset = unit * ai->unit_size;
1514 
1515                 for_each_possible_cpu(cpu)
1516                         if (group_map[cpu] == group)
1517                                 gi->cpu_map[gi->nr_units++] = cpu;
1518                 gi->nr_units = roundup(gi->nr_units, upa);
1519                 unit += gi->nr_units;
1520         }
1521         BUG_ON(unit != nr_units);
1522 
1523         return ai;
1524 }
1525 
1526 /**
1527  * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1528  * @lvl: loglevel
1529  * @ai: allocation info to dump
1530  *
1531  * Print out information about @ai using loglevel @lvl.
1532  */
1533 static void pcpu_dump_alloc_info(const char *lvl,
1534                                  const struct pcpu_alloc_info *ai)
1535 {
1536         int group_width = 1, cpu_width = 1, width;
1537         char empty_str[] = "--------";
1538         int alloc = 0, alloc_end = 0;
1539         int group, v;
1540         int upa, apl;   /* units per alloc, allocs per line */
1541 
1542         v = ai->nr_groups;
1543         while (v /= 10)
1544                 group_width++;
1545 
1546         v = num_possible_cpus();
1547         while (v /= 10)
1548                 cpu_width++;
1549         empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1550 
1551         upa = ai->alloc_size / ai->unit_size;
1552         width = upa * (cpu_width + 1) + group_width + 3;
1553         apl = rounddown_pow_of_two(max(60 / width, 1));
1554 
1555         printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1556                lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1557                ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1558 
1559         for (group = 0; group < ai->nr_groups; group++) {
1560                 const struct pcpu_group_info *gi = &ai->groups[group];
1561                 int unit = 0, unit_end = 0;
1562 
1563                 BUG_ON(gi->nr_units % upa);
1564                 for (alloc_end += gi->nr_units / upa;
1565                      alloc < alloc_end; alloc++) {
1566                         if (!(alloc % apl)) {
1567                                 printk("\n");
1568                                 printk("%spcpu-alloc: ", lvl);
1569                         }
1570                         printk("[%0*d] ", group_width, group);
1571 
1572                         for (unit_end += upa; unit < unit_end; unit++)
1573                                 if (gi->cpu_map[unit] != NR_CPUS)
1574                                         printk("%0*d ", cpu_width,
1575                                                gi->cpu_map[unit]);
1576                                 else
1577                                         printk("%s ", empty_str);
1578                 }
1579         }
1580         printk("\n");
1581 }
1582 
1583 /**
1584  * pcpu_setup_first_chunk - initialize the first percpu chunk
1585  * @ai: pcpu_alloc_info describing how to percpu area is shaped
1586  * @base_addr: mapped address
1587  *
1588  * Initialize the first percpu chunk which contains the kernel static
1589  * perpcu area.  This function is to be called from arch percpu area
1590  * setup path.
1591  *
1592  * @ai contains all information necessary to initialize the first
1593  * chunk and prime the dynamic percpu allocator.
1594  *
1595  * @ai->static_size is the size of static percpu area.
1596  *
1597  * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1598  * reserve after the static area in the first chunk.  This reserves
1599  * the first chunk such that it's available only through reserved
1600  * percpu allocation.  This is primarily used to serve module percpu
1601  * static areas on architectures where the addressing model has
1602  * limited offset range for symbol relocations to guarantee module
1603  * percpu symbols fall inside the relocatable range.
1604  *
1605  * @ai->dyn_size determines the number of bytes available for dynamic
1606  * allocation in the first chunk.  The area between @ai->static_size +
1607  * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1608  *
1609  * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1610  * and equal to or larger than @ai->static_size + @ai->reserved_size +
1611  * @ai->dyn_size.
1612  *
1613  * @ai->atom_size is the allocation atom size and used as alignment
1614  * for vm areas.
1615  *
1616  * @ai->alloc_size is the allocation size and always multiple of
1617  * @ai->atom_size.  This is larger than @ai->atom_size if
1618  * @ai->unit_size is larger than @ai->atom_size.
1619  *
1620  * @ai->nr_groups and @ai->groups describe virtual memory layout of
1621  * percpu areas.  Units which should be colocated are put into the
1622  * same group.  Dynamic VM areas will be allocated according to these
1623  * groupings.  If @ai->nr_groups is zero, a single group containing
1624  * all units is assumed.
1625  *
1626  * The caller should have mapped the first chunk at @base_addr and
1627  * copied static data to each unit.
1628  *
1629  * If the first chunk ends up with both reserved and dynamic areas, it
1630  * is served by two chunks - one to serve the core static and reserved
1631  * areas and the other for the dynamic area.  They share the same vm
1632  * and page map but uses different area allocation map to stay away
1633  * from each other.  The latter chunk is circulated in the chunk slots
1634  * and available for dynamic allocation like any other chunks.
1635  *
1636  * RETURNS:
1637  * 0 on success, -errno on failure.
1638  */
1639 int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1640                                   void *base_addr)
1641 {
1642         static char cpus_buf[4096] __initdata;
1643         static int smap[2], dmap[2];
1644         size_t dyn_size = ai->dyn_size;
1645         size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1646         struct pcpu_chunk *schunk, *dchunk = NULL;
1647         unsigned long *group_offsets;
1648         size_t *group_sizes;
1649         unsigned long *unit_off;
1650         unsigned int cpu;
1651         int *unit_map;
1652         int group, unit, i;
1653 
1654         cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1655 
1656 #define PCPU_SETUP_BUG_ON(cond) do {                                    \
1657         if (unlikely(cond)) {                                           \
1658                 pr_emerg("PERCPU: failed to initialize, %s", #cond);    \
1659                 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf);   \
1660                 pcpu_dump_alloc_info(KERN_EMERG, ai);                   \
1661                 BUG();                                                  \
1662         }                                                               \
1663 } while (0)
1664 
1665         /* sanity checks */
1666         BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1667                      ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
1668         PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1669         PCPU_SETUP_BUG_ON(!ai->static_size);
1670         PCPU_SETUP_BUG_ON(!base_addr);
1671         PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1672         PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1673         PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1674 
1675         /* process group information and build config tables accordingly */
1676         group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1677         group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
1678         unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
1679         unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
1680 
1681         for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1682                 unit_map[cpu] = UINT_MAX;
1683 
1684         pcpu_low_unit_cpu = NR_CPUS;
1685         pcpu_high_unit_cpu = NR_CPUS;
1686 
1687         for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1688                 const struct pcpu_group_info *gi = &ai->groups[group];
1689 
1690                 group_offsets[group] = gi->base_offset;
1691                 group_sizes[group] = gi->nr_units * ai->unit_size;
1692 
1693                 for (i = 0; i < gi->nr_units; i++) {
1694                         cpu = gi->cpu_map[i];
1695                         if (cpu == NR_CPUS)
1696                                 continue;
1697 
1698                         PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1699                         PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1700                         PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1701 
1702                         unit_map[cpu] = unit + i;
1703                         unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1704 
1705                         /* determine low/high unit_cpu */
1706                         if (pcpu_low_unit_cpu == NR_CPUS ||
1707                             unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1708                                 pcpu_low_unit_cpu = cpu;
1709                         if (pcpu_high_unit_cpu == NR_CPUS ||
1710                             unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1711                                 pcpu_high_unit_cpu = cpu;
1712                 }
1713         }
1714         pcpu_nr_units = unit;
1715 
1716         for_each_possible_cpu(cpu)
1717                 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1718 
1719         /* we're done parsing the input, undefine BUG macro and dump config */
1720 #undef PCPU_SETUP_BUG_ON
1721         pcpu_dump_alloc_info(KERN_INFO, ai);
1722 
1723         pcpu_nr_groups = ai->nr_groups;
1724         pcpu_group_offsets = group_offsets;
1725         pcpu_group_sizes = group_sizes;
1726         pcpu_unit_map = unit_map;
1727         pcpu_unit_offsets = unit_off;
1728 
1729         /* determine basic parameters */
1730         pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1731         pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1732         pcpu_atom_size = ai->atom_size;
1733         pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1734                 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1735 
1736         /*
1737          * Allocate chunk slots.  The additional last slot is for
1738          * empty chunks.
1739          */
1740         pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1741         pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1742         for (i = 0; i < pcpu_nr_slots; i++)
1743                 INIT_LIST_HEAD(&pcpu_slot[i]);
1744 
1745         /*
1746          * Initialize static chunk.  If reserved_size is zero, the
1747          * static chunk covers static area + dynamic allocation area
1748          * in the first chunk.  If reserved_size is not zero, it
1749          * covers static area + reserved area (mostly used for module
1750          * static percpu allocation).
1751          */
1752         schunk = alloc_bootmem(pcpu_chunk_struct_size);
1753         INIT_LIST_HEAD(&schunk->list);
1754         schunk->base_addr = base_addr;
1755         schunk->map = smap;
1756         schunk->map_alloc = ARRAY_SIZE(smap);
1757         schunk->immutable = true;
1758         bitmap_fill(schunk->populated, pcpu_unit_pages);
1759 
1760         if (ai->reserved_size) {
1761                 schunk->free_size = ai->reserved_size;
1762                 pcpu_reserved_chunk = schunk;
1763                 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1764         } else {
1765                 schunk->free_size = dyn_size;
1766                 dyn_size = 0;                   /* dynamic area covered */
1767         }
1768         schunk->contig_hint = schunk->free_size;
1769 
1770         schunk->map[schunk->map_used++] = -ai->static_size;
1771         if (schunk->free_size)
1772                 schunk->map[schunk->map_used++] = schunk->free_size;
1773 
1774         /* init dynamic chunk if necessary */
1775         if (dyn_size) {
1776                 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
1777                 INIT_LIST_HEAD(&dchunk->list);
1778                 dchunk->base_addr = base_addr;
1779                 dchunk->map = dmap;
1780                 dchunk->map_alloc = ARRAY_SIZE(dmap);
1781                 dchunk->immutable = true;
1782                 bitmap_fill(dchunk->populated, pcpu_unit_pages);
1783 
1784                 dchunk->contig_hint = dchunk->free_size = dyn_size;
1785                 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1786                 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1787         }
1788 
1789         /* link the first chunk in */
1790         pcpu_first_chunk = dchunk ?: schunk;
1791         pcpu_chunk_relocate(pcpu_first_chunk, -1);
1792 
1793         /* we're done */
1794         pcpu_base_addr = base_addr;
1795         return 0;
1796 }
1797 
1798 const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1799         [PCPU_FC_AUTO]  = "auto",
1800         [PCPU_FC_EMBED] = "embed",
1801         [PCPU_FC_PAGE]  = "page",
1802 };
1803 
1804 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1805 
1806 static int __init percpu_alloc_setup(char *str)
1807 {
1808         if (0)
1809                 /* nada */;
1810 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1811         else if (!strcmp(str, "embed"))
1812                 pcpu_chosen_fc = PCPU_FC_EMBED;
1813 #endif
1814 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1815         else if (!strcmp(str, "page"))
1816                 pcpu_chosen_fc = PCPU_FC_PAGE;
1817 #endif
1818         else
1819                 pr_warning("PERCPU: unknown allocator %s specified\n", str);
1820 
1821         return 0;
1822 }
1823 early_param("percpu_alloc", percpu_alloc_setup);
1824 
1825 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1826         !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1827 /**
1828  * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1829  * @reserved_size: the size of reserved percpu area in bytes
1830  * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1831  * @atom_size: allocation atom size
1832  * @cpu_distance_fn: callback to determine distance between cpus, optional
1833  * @alloc_fn: function to allocate percpu page
1834  * @free_fn: funtion to free percpu page
1835  *
1836  * This is a helper to ease setting up embedded first percpu chunk and
1837  * can be called where pcpu_setup_first_chunk() is expected.
1838  *
1839  * If this function is used to setup the first chunk, it is allocated
1840  * by calling @alloc_fn and used as-is without being mapped into
1841  * vmalloc area.  Allocations are always whole multiples of @atom_size
1842  * aligned to @atom_size.
1843  *
1844  * This enables the first chunk to piggy back on the linear physical
1845  * mapping which often uses larger page size.  Please note that this
1846  * can result in very sparse cpu->unit mapping on NUMA machines thus
1847  * requiring large vmalloc address space.  Don't use this allocator if
1848  * vmalloc space is not orders of magnitude larger than distances
1849  * between node memory addresses (ie. 32bit NUMA machines).
1850  *
1851  * When @dyn_size is positive, dynamic area might be larger than
1852  * specified to fill page alignment.  When @dyn_size is auto,
1853  * @dyn_size is just big enough to fill page alignment after static
1854  * and reserved areas.
1855  *
1856  * If the needed size is smaller than the minimum or specified unit
1857  * size, the leftover is returned using @free_fn.
1858  *
1859  * RETURNS:
1860  * 0 on success, -errno on failure.
1861  */
1862 int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size,
1863                                   size_t atom_size,
1864                                   pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1865                                   pcpu_fc_alloc_fn_t alloc_fn,
1866                                   pcpu_fc_free_fn_t free_fn)
1867 {
1868         void *base = (void *)ULONG_MAX;
1869         void **areas = NULL;
1870         struct pcpu_alloc_info *ai;
1871         size_t size_sum, areas_size, max_distance;
1872         int group, i, rc;
1873 
1874         ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1875                                    cpu_distance_fn);
1876         if (IS_ERR(ai))
1877                 return PTR_ERR(ai);
1878 
1879         size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1880         areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1881 
1882         areas = alloc_bootmem_nopanic(areas_size);
1883         if (!areas) {
1884                 rc = -ENOMEM;
1885                 goto out_free;
1886         }
1887 
1888         /* allocate, copy and determine base address */
1889         for (group = 0; group < ai->nr_groups; group++) {
1890                 struct pcpu_group_info *gi = &ai->groups[group];
1891                 unsigned int cpu = NR_CPUS;
1892                 void *ptr;
1893 
1894                 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1895                         cpu = gi->cpu_map[i];
1896                 BUG_ON(cpu == NR_CPUS);
1897 
1898                 /* allocate space for the whole group */
1899                 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1900                 if (!ptr) {
1901                         rc = -ENOMEM;
1902                         goto out_free_areas;
1903                 }
1904                 areas[group] = ptr;
1905 
1906                 base = min(ptr, base);
1907 
1908                 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1909                         if (gi->cpu_map[i] == NR_CPUS) {
1910                                 /* unused unit, free whole */
1911                                 free_fn(ptr, ai->unit_size);
1912                                 continue;
1913                         }
1914                         /* copy and return the unused part */
1915                         memcpy(ptr, __per_cpu_load, ai->static_size);
1916                         free_fn(ptr + size_sum, ai->unit_size - size_sum);
1917                 }
1918         }
1919 
1920         /* base address is now known, determine group base offsets */
1921         max_distance = 0;
1922         for (group = 0; group < ai->nr_groups; group++) {
1923                 ai->groups[group].base_offset = areas[group] - base;
1924                 max_distance = max_t(size_t, max_distance,
1925                                      ai->groups[group].base_offset);
1926         }
1927         max_distance += ai->unit_size;
1928 
1929         /* warn if maximum distance is further than 75% of vmalloc space */
1930         if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
1931                 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
1932                            "space 0x%lx\n",
1933                            max_distance, VMALLOC_END - VMALLOC_START);
1934 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1935                 /* and fail if we have fallback */
1936                 rc = -EINVAL;
1937                 goto out_free;
1938 #endif
1939         }
1940 
1941         pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
1942                 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1943                 ai->dyn_size, ai->unit_size);
1944 
1945         rc = pcpu_setup_first_chunk(ai, base);
1946         goto out_free;
1947 
1948 out_free_areas:
1949         for (group = 0; group < ai->nr_groups; group++)
1950                 free_fn(areas[group],
1951                         ai->groups[group].nr_units * ai->unit_size);
1952 out_free:
1953         pcpu_free_alloc_info(ai);
1954         if (areas)
1955                 free_bootmem(__pa(areas), areas_size);
1956         return rc;
1957 }
1958 #endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
1959           !CONFIG_HAVE_SETUP_PER_CPU_AREA */
1960 
1961 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1962 /**
1963  * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
1964  * @reserved_size: the size of reserved percpu area in bytes
1965  * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1966  * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1967  * @populate_pte_fn: function to populate pte
1968  *
1969  * This is a helper to ease setting up page-remapped first percpu
1970  * chunk and can be called where pcpu_setup_first_chunk() is expected.
1971  *
1972  * This is the basic allocator.  Static percpu area is allocated
1973  * page-by-page into vmalloc area.
1974  *
1975  * RETURNS:
1976  * 0 on success, -errno on failure.
1977  */
1978 int __init pcpu_page_first_chunk(size_t reserved_size,
1979                                  pcpu_fc_alloc_fn_t alloc_fn,
1980                                  pcpu_fc_free_fn_t free_fn,
1981                                  pcpu_fc_populate_pte_fn_t populate_pte_fn)
1982 {
1983         static struct vm_struct vm;
1984         struct pcpu_alloc_info *ai;
1985         char psize_str[16];
1986         int unit_pages;
1987         size_t pages_size;
1988         struct page **pages;
1989         int unit, i, j, rc;
1990 
1991         snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1992 
1993         ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL);
1994         if (IS_ERR(ai))
1995                 return PTR_ERR(ai);
1996         BUG_ON(ai->nr_groups != 1);
1997         BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1998 
1999         unit_pages = ai->unit_size >> PAGE_SHIFT;
2000 
2001         /* unaligned allocations can't be freed, round up to page size */
2002         pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2003                                sizeof(pages[0]));
2004         pages = alloc_bootmem(pages_size);
2005 
2006         /* allocate pages */
2007         j = 0;
2008         for (unit = 0; unit < num_possible_cpus(); unit++)
2009                 for (i = 0; i < unit_pages; i++) {
2010                         unsigned int cpu = ai->groups[0].cpu_map[unit];
2011                         void *ptr;
2012 
2013                         ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2014                         if (!ptr) {
2015                                 pr_warning("PERCPU: failed to allocate %s page "
2016                                            "for cpu%u\n", psize_str, cpu);
2017                                 goto enomem;
2018                         }
2019                         pages[j++] = virt_to_page(ptr);
2020                 }
2021 
2022         /* allocate vm area, map the pages and copy static data */
2023         vm.flags = VM_ALLOC;
2024         vm.size = num_possible_cpus() * ai->unit_size;
2025         vm_area_register_early(&vm, PAGE_SIZE);
2026 
2027         for (unit = 0; unit < num_possible_cpus(); unit++) {
2028                 unsigned long unit_addr =
2029                         (unsigned long)vm.addr + unit * ai->unit_size;
2030 
2031                 for (i = 0; i < unit_pages; i++)
2032                         populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2033 
2034                 /* pte already populated, the following shouldn't fail */
2035                 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2036                                       unit_pages);
2037                 if (rc < 0)
2038                         panic("failed to map percpu area, err=%d\n", rc);
2039 
2040                 /*
2041                  * FIXME: Archs with virtual cache should flush local
2042                  * cache for the linear mapping here - something
2043                  * equivalent to flush_cache_vmap() on the local cpu.
2044                  * flush_cache_vmap() can't be used as most supporting
2045                  * data structures are not set up yet.
2046                  */
2047 
2048                 /* copy static data */
2049                 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2050         }
2051 
2052         /* we're ready, commit */
2053         pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
2054                 unit_pages, psize_str, vm.addr, ai->static_size,
2055                 ai->reserved_size, ai->dyn_size);
2056 
2057         rc = pcpu_setup_first_chunk(ai, vm.addr);
2058         goto out_free_ar;
2059 
2060 enomem:
2061         while (--j >= 0)
2062                 free_fn(page_address(pages[j]), PAGE_SIZE);
2063         rc = -ENOMEM;
2064 out_free_ar:
2065         free_bootmem(__pa(pages), pages_size);
2066         pcpu_free_alloc_info(ai);
2067         return rc;
2068 }
2069 #endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
2070 
2071 /*
2072  * Generic percpu area setup.
2073  *
2074  * The embedding helper is used because its behavior closely resembles
2075  * the original non-dynamic generic percpu area setup.  This is
2076  * important because many archs have addressing restrictions and might
2077  * fail if the percpu area is located far away from the previous
2078  * location.  As an added bonus, in non-NUMA cases, embedding is
2079  * generally a good idea TLB-wise because percpu area can piggy back
2080  * on the physical linear memory mapping which uses large page
2081  * mappings on applicable archs.
2082  */
2083 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2084 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2085 EXPORT_SYMBOL(__per_cpu_offset);
2086 
2087 static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2088                                        size_t align)
2089 {
2090         return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
2091 }
2092 
2093 static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2094 {
2095         free_bootmem(__pa(ptr), size);
2096 }
2097 
2098 void __init setup_per_cpu_areas(void)
2099 {
2100         unsigned long delta;
2101         unsigned int cpu;
2102         int rc;
2103 
2104         /*
2105          * Always reserve area for module percpu variables.  That's
2106          * what the legacy allocator did.
2107          */
2108         rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2109                                     PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2110                                     pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2111         if (rc < 0)
2112                 panic("Failed to initialized percpu areas.");
2113 
2114         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2115         for_each_possible_cpu(cpu)
2116                 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2117 }
2118 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2119 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp