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, 0);
  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  * @gfp: allocation flags passed to the underlying allocator
  77  *
  78  * Allocate pages [@page_start,@page_end) into @pages for all units.
  79  * The allocation is for @chunk.  Percpu core doesn't care about the
  80  * content of @pages and will pass it verbatim to pcpu_map_pages().
  81  */
  82 static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
  83                             struct page **pages, int page_start, int page_end,
  84                             gfp_t gfp)
  85 {
  86         unsigned int cpu, tcpu;
  87         int i;
  88
  89         gfp |= GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
  90
  91         for_each_possible_cpu(cpu) {
  92                 for (i = page_start; i < page_end; i++) {
  93                         struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
  94
  95                         *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
  96                         if (!*pagep)
  97                                 goto err;
  98                 }
  99         }
 100         return 0;
 101
 102 err:
 103         while (--i >= page_start)
 104                 __free_page(pages[pcpu_page_idx(cpu, i)]);
 105
 106         for_each_possible_cpu(tcpu) {
 107                 if (tcpu == cpu)
 108                         break;
 109                 for (i = page_start; i < page_end; i++)
 110                         __free_page(pages[pcpu_page_idx(tcpu, i)]);
 111         }
 112         return -ENOMEM;
 113 }
 114
 115 /**
 116  * pcpu_pre_unmap_flush - flush cache prior to unmapping
 117  * @chunk: chunk the regions to be flushed belongs to
 118  * @page_start: page index of the first page to be flushed
 119  * @page_end: page index of the last page to be flushed + 1
 120  *
 121  * Pages in [@page_start,@page_end) of @chunk are about to be
 122  * unmapped.  Flush cache.  As each flushing trial can be very
 123  * expensive, issue flush on the whole region at once rather than
 124  * doing it for each cpu.  This could be an overkill but is more
 125  * scalable.
 126  */
 127 static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
 128                                  int page_start, int page_end)
 129 {
 130         flush_cache_vunmap(
 131                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 132                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 133 }
 134
 135 static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
 136 {
 137         unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
 138 }
 139
 140 /**
 141  * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
 142  * @chunk: chunk of interest
 143  * @pages: pages array which can be used to pass information to free
 144  * @page_start: page index of the first page to unmap
 145  * @page_end: page index of the last page to unmap + 1
 146  *
 147  * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
 148  * Corresponding elements in @pages were cleared by the caller and can
 149  * be used to carry information to pcpu_free_pages() which will be
 150  * called after all unmaps are finished.  The caller should call
 151  * proper pre/post flush functions.
 152  */
 153 static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
 154                              struct page **pages, int page_start, int page_end)
 155 {
 156         unsigned int cpu;
 157         int i;
 158
 159         for_each_possible_cpu(cpu) {
 160                 for (i = page_start; i < page_end; i++) {
 161                         struct page *page;
 162
 163                         page = pcpu_chunk_page(chunk, cpu, i);
 164                         WARN_ON(!page);
 165                         pages[pcpu_page_idx(cpu, i)] = page;
 166                 }
 167                 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
 168                                    page_end - page_start);
 169         }
 170 }
 171
 172 /**
 173  * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
 174  * @chunk: pcpu_chunk the regions to be flushed belong to
 175  * @page_start: page index of the first page to be flushed
 176  * @page_end: page index of the last page to be flushed + 1
 177  *
 178  * Pages [@page_start,@page_end) of @chunk have been unmapped.  Flush
 179  * TLB for the regions.  This can be skipped if the area is to be
 180  * returned to vmalloc as vmalloc will handle TLB flushing lazily.
 181  *
 182  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
 183  * for the whole region.
 184  */
 185 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
 186                                       int page_start, int page_end)
 187 {
 188         flush_tlb_kernel_range(
 189                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 190                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 191 }
 192
 193 static int __pcpu_map_pages(unsigned long addr, struct page **pages,
 194                             int nr_pages)
 195 {
 196         return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
 197                                         PAGE_KERNEL, pages);
 198 }
 199
 200 /**
 201  * pcpu_map_pages - map pages into a pcpu_chunk
 202  * @chunk: chunk of interest
 203  * @pages: pages array containing pages to be mapped
 204  * @page_start: page index of the first page to map
 205  * @page_end: page index of the last page to map + 1
 206  *
 207  * For each cpu, map pages [@page_start,@page_end) into @chunk.  The
 208  * caller is responsible for calling pcpu_post_map_flush() after all
 209  * mappings are complete.
 210  *
 211  * This function is responsible for setting up whatever is necessary for
 212  * reverse lookup (addr -> chunk).
 213  */
 214 static int pcpu_map_pages(struct pcpu_chunk *chunk,
 215                           struct page **pages, int page_start, int page_end)
 216 {
 217         unsigned int cpu, tcpu;
 218         int i, err;
 219
 220         for_each_possible_cpu(cpu) {
 221                 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
 222                                        &pages[pcpu_page_idx(cpu, page_start)],
 223                                        page_end - page_start);
 224                 if (err < 0)
 225                         goto err;
 226
 227                 for (i = page_start; i < page_end; i++)
 228                         pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
 229                                             chunk);
 230         }
 231         return 0;
 232 err:
 233         for_each_possible_cpu(tcpu) {
 234                 if (tcpu == cpu)
 235                         break;
 236                 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
 237                                    page_end - page_start);
 238         }
 239         pcpu_post_unmap_tlb_flush(chunk, page_start, page_end);
 240         return err;
 241 }
 242
 243 /**
 244  * pcpu_post_map_flush - flush cache after mapping
 245  * @chunk: pcpu_chunk the regions to be flushed belong to
 246  * @page_start: page index of the first page to be flushed
 247  * @page_end: page index of the last page to be flushed + 1
 248  *
 249  * Pages [@page_start,@page_end) of @chunk have been mapped.  Flush
 250  * cache.
 251  *
 252  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
 253  * for the whole region.
 254  */
 255 static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
 256                                 int page_start, int page_end)
 257 {
 258         flush_cache_vmap(
 259                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 260                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 261 }
 262
 263 /**
 264  * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
 265  * @chunk: chunk of interest
 266  * @page_start: the start page
 267  * @page_end: the end page
 268  * @gfp: allocation flags passed to the underlying memory allocator
 269  *
 270  * For each cpu, populate and map pages [@page_start,@page_end) into
 271  * @chunk.
 272  *
 273  * CONTEXT:
 274  * pcpu_alloc_mutex, does GFP_KERNEL allocation.
 275  */
 276 static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
 277                                int page_start, int page_end, gfp_t gfp)
 278 {
 279         struct page **pages;
 280
 281         pages = pcpu_get_pages();
 282         if (!pages)
 283                 return -ENOMEM;
 284
 285         if (pcpu_alloc_pages(chunk, pages, page_start, page_end, gfp))
 286                 return -ENOMEM;
 287
 288         if (pcpu_map_pages(chunk, pages, page_start, page_end)) {
 289                 pcpu_free_pages(chunk, pages, page_start, page_end);
 290                 return -ENOMEM;
 291         }
 292         pcpu_post_map_flush(chunk, page_start, page_end);
 293
 294         return 0;
 295 }
 296
 297 /**
 298  * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
 299  * @chunk: chunk to depopulate
 300  * @page_start: the start page
 301  * @page_end: the end page
 302  *
 303  * For each cpu, depopulate and unmap pages [@page_start,@page_end)
 304  * from @chunk.
 305  *
 306  * CONTEXT:
 307  * pcpu_alloc_mutex.
 308  */
 309 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
 310                                   int page_start, int page_end)
 311 {
 312         struct page **pages;
 313
 314         /*
 315          * If control reaches here, there must have been at least one
 316          * successful population attempt so the temp pages array must
 317          * be available now.
 318          */
 319         pages = pcpu_get_pages();
 320         BUG_ON(!pages);
 321
 322         /* unmap and free */
 323         pcpu_pre_unmap_flush(chunk, page_start, page_end);
 324
 325         pcpu_unmap_pages(chunk, pages, page_start, page_end);
 326
 327         /* no need to flush tlb, vmalloc will handle it lazily */
 328
 329         pcpu_free_pages(chunk, pages, page_start, page_end);
 330 }
 331
 332 static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp)
 333 {
 334         struct pcpu_chunk *chunk;
 335         struct vm_struct **vms;
 336
 337         chunk = pcpu_alloc_chunk(gfp);
 338         if (!chunk)
 339                 return NULL;
 340
 341         vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
 342                                 pcpu_nr_groups, pcpu_atom_size);
 343         if (!vms) {
 344                 pcpu_free_chunk(chunk);
 345                 return NULL;
 346         }
 347
 348         chunk->data = vms;
 349         chunk->base_addr = vms[0]->addr - pcpu_group_offsets[0];
 350
 351         pcpu_stats_chunk_alloc();
 352         trace_percpu_create_chunk(chunk->base_addr);
 353
 354         return chunk;
 355 }
 356
 357 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk)
 358 {
 359         if (!chunk)
 360                 return;
 361
 362         pcpu_stats_chunk_dealloc();
 363         trace_percpu_destroy_chunk(chunk->base_addr);
 364
 365         if (chunk->data)
 366                 pcpu_free_vm_areas(chunk->data, pcpu_nr_groups);
 367         pcpu_free_chunk(chunk);
 368 }
 369
 370 static struct page *pcpu_addr_to_page(void *addr)
 371 {
 372         return vmalloc_to_page(addr);
 373 }
 374
 375 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai)
 376 {
 377         /* no extra restriction */
 378         return 0;
 379 }