mm/percpu-vm.c

   1 /*
   2  * mm/percpu-vm.c - vmalloc area based chunk allocation
   3  *
   4  * Copyright (C) 2010           SUSE Linux Products GmbH
   5  * Copyright (C) 2010           Tejun Heo <tj@kernel.org>
   6  *
   7  * This file is released under the GPLv2.
   8  *
   9  * Chunks are mapped into vmalloc areas and populated page by page.
  10  * This is the default chunk allocator.
  11  */
  12
  13 static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
  14                                     unsigned int cpu, int page_idx)
  15 {
  16         /* must not be used on pre-mapped chunk */
  17         WARN_ON(chunk->immutable);
  18
  19         return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
  20 }
  21
  22 /**
  23  * pcpu_get_pages - get temp pages array
  24  *
  25  * Returns pointer to array of pointers to struct page which can be indexed
  26  * with pcpu_page_idx().  Note that there is only one array and accesses
  27  * should be serialized by pcpu_alloc_mutex.
  28  *
  29  * RETURNS:
  30  * Pointer to temp pages array on success.
  31  */
  32 static struct page **pcpu_get_pages(void)
  33 {
  34         static struct page **pages;
  35         size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
  36
  37         lockdep_assert_held(&pcpu_alloc_mutex);
  38
  39         if (!pages)
  40                 pages = pcpu_mem_zalloc(pages_size);
  41         return pages;
  42 }
  43
  44 /**
  45  * pcpu_free_pages - free pages which were allocated for @chunk
  46  * @chunk: chunk pages were allocated for
  47  * @pages: array of pages to be freed, indexed by pcpu_page_idx()
  48  * @page_start: page index of the first page to be freed
  49  * @page_end: page index of the last page to be freed + 1
  50  *
  51  * Free pages [@page_start and @page_end) in @pages for all units.
  52  * The pages were allocated for @chunk.
  53  */
  54 static void pcpu_free_pages(struct pcpu_chunk *chunk,
  55                             struct page **pages, int page_start, int page_end)
  56 {
  57         unsigned int cpu;
  58         int i;
  59
  60         for_each_possible_cpu(cpu) {
  61                 for (i = page_start; i < page_end; i++) {
  62                         struct page *page = pages[pcpu_page_idx(cpu, i)];
  63
  64                         if (page)
  65                                 __free_page(page);
  66                 }
  67         }
  68 }
  69
  70 /**
  71  * pcpu_alloc_pages - allocates pages for @chunk
  72  * @chunk: target chunk
  73  * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
  74  * @page_start: page index of the first page to be allocated
  75  * @page_end: page index of the last page to be allocated + 1
  76  *
  77  * Allocate pages [@page_start,@page_end) into @pages for all units.
  78  * The allocation is for @chunk.  Percpu core doesn't care about the
  79  * content of @pages and will pass it verbatim to pcpu_map_pages().
  80  */
  81 static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
  82                             struct page **pages, int page_start, int page_end)
  83 {
  84         const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
  85         unsigned int cpu, tcpu;
  86         int i;
  87
  88         for_each_possible_cpu(cpu) {
  89                 for (i = page_start; i < page_end; i++) {
  90                         struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
  91
  92                         *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
  93                         if (!*pagep)
  94                                 goto err;
  95                 }
  96         }
  97         return 0;
  98
  99 err:
 100         while (--i >= page_start)
 101                 __free_page(pages[pcpu_page_idx(cpu, i)]);
 102
 103         for_each_possible_cpu(tcpu) {
 104                 if (tcpu == cpu)
 105                         break;
 106                 for (i = page_start; i < page_end; i++)
 107                         __free_page(pages[pcpu_page_idx(tcpu, i)]);
 108         }
 109         return -ENOMEM;
 110 }
 111
 112 /**
 113  * pcpu_pre_unmap_flush - flush cache prior to unmapping
 114  * @chunk: chunk the regions to be flushed belongs to
 115  * @page_start: page index of the first page to be flushed
 116  * @page_end: page index of the last page to be flushed + 1
 117  *
 118  * Pages in [@page_start,@page_end) of @chunk are about to be
 119  * unmapped.  Flush cache.  As each flushing trial can be very
 120  * expensive, issue flush on the whole region at once rather than
 121  * doing it for each cpu.  This could be an overkill but is more
 122  * scalable.
 123  */
 124 static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
 125                                  int page_start, int page_end)
 126 {
 127         flush_cache_vunmap(
 128                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 129                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 130 }
 131
 132 static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
 133 {
 134         unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
 135 }
 136
 137 /**
 138  * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
 139  * @chunk: chunk of interest
 140  * @pages: pages array which can be used to pass information to free
 141  * @page_start: page index of the first page to unmap
 142  * @page_end: page index of the last page to unmap + 1
 143  *
 144  * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
 145  * Corresponding elements in @pages were cleared by the caller and can
 146  * be used to carry information to pcpu_free_pages() which will be
 147  * called after all unmaps are finished.  The caller should call
 148  * proper pre/post flush functions.
 149  */
 150 static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
 151                              struct page **pages, int page_start, int page_end)
 152 {
 153         unsigned int cpu;
 154         int i;
 155
 156         for_each_possible_cpu(cpu) {
 157                 for (i = page_start; i < page_end; i++) {
 158                         struct page *page;
 159
 160                         page = pcpu_chunk_page(chunk, cpu, i);
 161                         WARN_ON(!page);
 162                         pages[pcpu_page_idx(cpu, i)] = page;
 163                 }
 164                 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
 165                                    page_end - page_start);
 166         }
 167 }
 168
 169 /**
 170  * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
 171  * @chunk: pcpu_chunk the regions to be flushed belong to
 172  * @page_start: page index of the first page to be flushed
 173  * @page_end: page index of the last page to be flushed + 1
 174  *
 175  * Pages [@page_start,@page_end) of @chunk have been unmapped.  Flush
 176  * TLB for the regions.  This can be skipped if the area is to be
 177  * returned to vmalloc as vmalloc will handle TLB flushing lazily.
 178  *
 179  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
 180  * for the whole region.
 181  */
 182 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
 183                                       int page_start, int page_end)
 184 {
 185         flush_tlb_kernel_range(
 186                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 187                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 188 }
 189
 190 static int __pcpu_map_pages(unsigned long addr, struct page **pages,
 191                             int nr_pages)
 192 {
 193         return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
 194                                         PAGE_KERNEL, pages);
 195 }
 196
 197 /**
 198  * pcpu_map_pages - map pages into a pcpu_chunk
 199  * @chunk: chunk of interest
 200  * @pages: pages array containing pages to be mapped
 201  * @page_start: page index of the first page to map
 202  * @page_end: page index of the last page to map + 1
 203  *
 204  * For each cpu, map pages [@page_start,@page_end) into @chunk.  The
 205  * caller is responsible for calling pcpu_post_map_flush() after all
 206  * mappings are complete.
 207  *
 208  * This function is responsible for setting up whatever is necessary for
 209  * reverse lookup (addr -> chunk).
 210  */
 211 static int pcpu_map_pages(struct pcpu_chunk *chunk,
 212                           struct page **pages, int page_start, int page_end)
 213 {
 214         unsigned int cpu, tcpu;
 215         int i, err;
 216
 217         for_each_possible_cpu(cpu) {
 218                 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
 219                                        &pages[pcpu_page_idx(cpu, page_start)],
 220                                        page_end - page_start);
 221                 if (err < 0)
 222                         goto err;
 223
 224                 for (i = page_start; i < page_end; i++)
 225                         pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
 226                                             chunk);
 227         }
 228         return 0;
 229 err:
 230         for_each_possible_cpu(tcpu) {
 231                 if (tcpu == cpu)
 232                         break;
 233                 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
 234                                    page_end - page_start);
 235         }
 236         pcpu_post_unmap_tlb_flush(chunk, page_start, page_end);
 237         return err;
 238 }
 239
 240 /**
 241  * pcpu_post_map_flush - flush cache after mapping
 242  * @chunk: pcpu_chunk the regions to be flushed belong to
 243  * @page_start: page index of the first page to be flushed
 244  * @page_end: page index of the last page to be flushed + 1
 245  *
 246  * Pages [@page_start,@page_end) of @chunk have been mapped.  Flush
 247  * cache.
 248  *
 249  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
 250  * for the whole region.
 251  */
 252 static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
 253                                 int page_start, int page_end)
 254 {
 255         flush_cache_vmap(
 256                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 257                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 258 }
 259
 260 /**
 261  * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
 262  * @chunk: chunk of interest
 263  * @page_start: the start page
 264  * @page_end: the end page
 265  *
 266  * For each cpu, populate and map pages [@page_start,@page_end) into
 267  * @chunk.
 268  *
 269  * CONTEXT:
 270  * pcpu_alloc_mutex, does GFP_KERNEL allocation.
 271  */
 272 static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
 273                                int page_start, int page_end)
 274 {
 275         struct page **pages;
 276
 277         pages = pcpu_get_pages();
 278         if (!pages)
 279                 return -ENOMEM;
 280
 281         if (pcpu_alloc_pages(chunk, pages, page_start, page_end))
 282                 return -ENOMEM;
 283
 284         if (pcpu_map_pages(chunk, pages, page_start, page_end)) {
 285                 pcpu_free_pages(chunk, pages, page_start, page_end);
 286                 return -ENOMEM;
 287         }
 288         pcpu_post_map_flush(chunk, page_start, page_end);
 289
 290         return 0;
 291 }
 292
 293 /**
 294  * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
 295  * @chunk: chunk to depopulate
 296  * @page_start: the start page
 297  * @page_end: the end page
 298  *
 299  * For each cpu, depopulate and unmap pages [@page_start,@page_end)
 300  * from @chunk.
 301  *
 302  * CONTEXT:
 303  * pcpu_alloc_mutex.
 304  */
 305 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
 306                                   int page_start, int page_end)
 307 {
 308         struct page **pages;
 309
 310         /*
 311          * If control reaches here, there must have been at least one
 312          * successful population attempt so the temp pages array must
 313          * be available now.
 314          */
 315         pages = pcpu_get_pages();
 316         BUG_ON(!pages);
 317
 318         /* unmap and free */
 319         pcpu_pre_unmap_flush(chunk, page_start, page_end);
 320
 321         pcpu_unmap_pages(chunk, pages, page_start, page_end);
 322
 323         /* no need to flush tlb, vmalloc will handle it lazily */
 324
 325         pcpu_free_pages(chunk, pages, page_start, page_end);
 326 }
 327
 328 static struct pcpu_chunk *pcpu_create_chunk(void)
 329 {
 330         struct pcpu_chunk *chunk;
 331         struct vm_struct **vms;
 332
 333         chunk = pcpu_alloc_chunk();
 334         if (!chunk)
 335                 return NULL;
 336
 337         vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
 338                                 pcpu_nr_groups, pcpu_atom_size);
 339         if (!vms) {
 340                 pcpu_free_chunk(chunk);
 341                 return NULL;
 342         }
 343
 344         chunk->data = vms;
 345         chunk->base_addr = vms[0]->addr - pcpu_group_offsets[0];
 346
 347         pcpu_stats_chunk_alloc();
 348         trace_percpu_create_chunk(chunk->base_addr);
 349
 350         return chunk;
 351 }
 352
 353 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk)
 354 {
 355         if (!chunk)
 356                 return;
 357
 358         pcpu_stats_chunk_dealloc();
 359         trace_percpu_destroy_chunk(chunk->base_addr);
 360
 361         if (chunk->data)
 362                 pcpu_free_vm_areas(chunk->data, pcpu_nr_groups);
 363         pcpu_free_chunk(chunk);
 364 }
 365
 366 static struct page *pcpu_addr_to_page(void *addr)
 367 {
 368         return vmalloc_to_page(addr);
 369 }
 370
 371 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai)
 372 {
 373         /* no extra restriction */
 374         return 0;
 375 }