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_and_bitmap - get temp pages array and bitmap
  24  * @chunk: chunk of interest
  25  * @bitmapp: output parameter for bitmap
  26  * @may_alloc: may allocate the array
  27  *
  28  * Returns pointer to array of pointers to struct page and bitmap,
  29  * both of which can be indexed with pcpu_page_idx().  The returned
  30  * array is cleared to zero and *@bitmapp is copied from
  31  * @chunk->populated.  Note that there is only one array and bitmap
  32  * and access exclusion is the caller's responsibility.
  33  *
  34  * CONTEXT:
  35  * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
  36  * Otherwise, don't care.
  37  *
  38  * RETURNS:
  39  * Pointer to temp pages array on success, NULL on failure.
  40  */
  41 static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
  42                                                unsigned long **bitmapp,
  43                                                bool may_alloc)
  44 {
  45         static struct page **pages;
  46         static unsigned long *bitmap;
  47         size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
  48         size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
  49                              sizeof(unsigned long);
  50
  51         if (!pages || !bitmap) {
  52                 if (may_alloc && !pages)
  53                         pages = pcpu_mem_zalloc(pages_size);
  54                 if (may_alloc && !bitmap)
  55                         bitmap = pcpu_mem_zalloc(bitmap_size);
  56                 if (!pages || !bitmap)
  57                         return NULL;
  58         }
  59
  60         bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
  61
  62         *bitmapp = bitmap;
  63         return pages;
  64 }
  65
  66 /**
  67  * pcpu_free_pages - free pages which were allocated for @chunk
  68  * @chunk: chunk pages were allocated for
  69  * @pages: array of pages to be freed, indexed by pcpu_page_idx()
  70  * @populated: populated bitmap
  71  * @page_start: page index of the first page to be freed
  72  * @page_end: page index of the last page to be freed + 1
  73  *
  74  * Free pages [@page_start and @page_end) in @pages for all units.
  75  * The pages were allocated for @chunk.
  76  */
  77 static void pcpu_free_pages(struct pcpu_chunk *chunk,
  78                             struct page **pages, unsigned long *populated,
  79                             int page_start, int page_end)
  80 {
  81         unsigned int cpu;
  82         int i;
  83
  84         for_each_possible_cpu(cpu) {
  85                 for (i = page_start; i < page_end; i++) {
  86                         struct page *page = pages[pcpu_page_idx(cpu, i)];
  87
  88                         if (page)
  89                                 __free_page(page);
  90                 }
  91         }
  92 }
  93
  94 /**
  95  * pcpu_alloc_pages - allocates pages for @chunk
  96  * @chunk: target chunk
  97  * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
  98  * @populated: populated bitmap
  99  * @page_start: page index of the first page to be allocated
 100  * @page_end: page index of the last page to be allocated + 1
 101  *
 102  * Allocate pages [@page_start,@page_end) into @pages for all units.
 103  * The allocation is for @chunk.  Percpu core doesn't care about the
 104  * content of @pages and will pass it verbatim to pcpu_map_pages().
 105  */
 106 static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
 107                             struct page **pages, unsigned long *populated,
 108                             int page_start, int page_end)
 109 {
 110         const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
 111         unsigned int cpu, tcpu;
 112         int i;
 113
 114         for_each_possible_cpu(cpu) {
 115                 for (i = page_start; i < page_end; i++) {
 116                         struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
 117
 118                         *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
 119                         if (!*pagep)
 120                                 goto err;
 121                 }
 122         }
 123         return 0;
 124
 125 err:
 126         while (--i >= page_start)
 127                 __free_page(pages[pcpu_page_idx(cpu, i)]);
 128
 129         for_each_possible_cpu(tcpu) {
 130                 if (tcpu == cpu)
 131                         break;
 132                 for (i = page_start; i < page_end; i++)
 133                         __free_page(pages[pcpu_page_idx(tcpu, i)]);
 134         }
 135         return -ENOMEM;
 136 }
 137
 138 /**
 139  * pcpu_pre_unmap_flush - flush cache prior to unmapping
 140  * @chunk: chunk the regions to be flushed belongs to
 141  * @page_start: page index of the first page to be flushed
 142  * @page_end: page index of the last page to be flushed + 1
 143  *
 144  * Pages in [@page_start,@page_end) of @chunk are about to be
 145  * unmapped.  Flush cache.  As each flushing trial can be very
 146  * expensive, issue flush on the whole region at once rather than
 147  * doing it for each cpu.  This could be an overkill but is more
 148  * scalable.
 149  */
 150 static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
 151                                  int page_start, int page_end)
 152 {
 153         flush_cache_vunmap(
 154                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 155                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 156 }
 157
 158 static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
 159 {
 160         unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
 161 }
 162
 163 /**
 164  * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
 165  * @chunk: chunk of interest
 166  * @pages: pages array which can be used to pass information to free
 167  * @populated: populated bitmap
 168  * @page_start: page index of the first page to unmap
 169  * @page_end: page index of the last page to unmap + 1
 170  *
 171  * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
 172  * Corresponding elements in @pages were cleared by the caller and can
 173  * be used to carry information to pcpu_free_pages() which will be
 174  * called after all unmaps are finished.  The caller should call
 175  * proper pre/post flush functions.
 176  */
 177 static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
 178                              struct page **pages, unsigned long *populated,
 179                              int page_start, int page_end)
 180 {
 181         unsigned int cpu;
 182         int i;
 183
 184         for_each_possible_cpu(cpu) {
 185                 for (i = page_start; i < page_end; i++) {
 186                         struct page *page;
 187
 188                         page = pcpu_chunk_page(chunk, cpu, i);
 189                         WARN_ON(!page);
 190                         pages[pcpu_page_idx(cpu, i)] = page;
 191                 }
 192                 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
 193                                    page_end - page_start);
 194         }
 195
 196         for (i = page_start; i < page_end; i++)
 197                 __clear_bit(i, populated);
 198 }
 199
 200 /**
 201  * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
 202  * @chunk: pcpu_chunk the regions to be flushed belong to
 203  * @page_start: page index of the first page to be flushed
 204  * @page_end: page index of the last page to be flushed + 1
 205  *
 206  * Pages [@page_start,@page_end) of @chunk have been unmapped.  Flush
 207  * TLB for the regions.  This can be skipped if the area is to be
 208  * returned to vmalloc as vmalloc will handle TLB flushing lazily.
 209  *
 210  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
 211  * for the whole region.
 212  */
 213 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
 214                                       int page_start, int page_end)
 215 {
 216         flush_tlb_kernel_range(
 217                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 218                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 219 }
 220
 221 static int __pcpu_map_pages(unsigned long addr, struct page **pages,
 222                             int nr_pages)
 223 {
 224         return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
 225                                         PAGE_KERNEL, pages);
 226 }
 227
 228 /**
 229  * pcpu_map_pages - map pages into a pcpu_chunk
 230  * @chunk: chunk of interest
 231  * @pages: pages array containing pages to be mapped
 232  * @populated: populated bitmap
 233  * @page_start: page index of the first page to map
 234  * @page_end: page index of the last page to map + 1
 235  *
 236  * For each cpu, map pages [@page_start,@page_end) into @chunk.  The
 237  * caller is responsible for calling pcpu_post_map_flush() after all
 238  * mappings are complete.
 239  *
 240  * This function is responsible for setting corresponding bits in
 241  * @chunk->populated bitmap and whatever is necessary for reverse
 242  * lookup (addr -> chunk).
 243  */
 244 static int pcpu_map_pages(struct pcpu_chunk *chunk,
 245                           struct page **pages, unsigned long *populated,
 246                           int page_start, int page_end)
 247 {
 248         unsigned int cpu, tcpu;
 249         int i, err;
 250
 251         for_each_possible_cpu(cpu) {
 252                 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
 253                                        &pages[pcpu_page_idx(cpu, page_start)],
 254                                        page_end - page_start);
 255                 if (err < 0)
 256                         goto err;
 257         }
 258
 259         /* mapping successful, link chunk and mark populated */
 260         for (i = page_start; i < page_end; i++) {
 261                 for_each_possible_cpu(cpu)
 262                         pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
 263                                             chunk);
 264                 __set_bit(i, populated);
 265         }
 266
 267         return 0;
 268
 269 err:
 270         for_each_possible_cpu(tcpu) {
 271                 if (tcpu == cpu)
 272                         break;
 273                 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
 274                                    page_end - page_start);
 275         }
 276         pcpu_post_unmap_tlb_flush(chunk, page_start, page_end);
 277         return err;
 278 }
 279
 280 /**
 281  * pcpu_post_map_flush - flush cache after mapping
 282  * @chunk: pcpu_chunk the regions to be flushed belong to
 283  * @page_start: page index of the first page to be flushed
 284  * @page_end: page index of the last page to be flushed + 1
 285  *
 286  * Pages [@page_start,@page_end) of @chunk have been mapped.  Flush
 287  * cache.
 288  *
 289  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
 290  * for the whole region.
 291  */
 292 static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
 293                                 int page_start, int page_end)
 294 {
 295         flush_cache_vmap(
 296                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 297                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 298 }
 299
 300 /**
 301  * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
 302  * @chunk: chunk of interest
 303  * @off: offset to the area to populate
 304  * @size: size of the area to populate in bytes
 305  *
 306  * For each cpu, populate and map pages [@page_start,@page_end) into
 307  * @chunk.  The area is cleared on return.
 308  *
 309  * CONTEXT:
 310  * pcpu_alloc_mutex, does GFP_KERNEL allocation.
 311  */
 312 static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
 313 {
 314         int page_start = PFN_DOWN(off);
 315         int page_end = PFN_UP(off + size);
 316         int free_end = page_start, unmap_end = page_start;
 317         struct page **pages;
 318         unsigned long *populated;
 319         unsigned int cpu;
 320         int rs, re, rc;
 321
 322         /* quick path, check whether all pages are already there */
 323         rs = page_start;
 324         pcpu_next_pop(chunk, &rs, &re, page_end);
 325         if (rs == page_start && re == page_end)
 326                 goto clear;
 327
 328         /* need to allocate and map pages, this chunk can't be immutable */
 329         WARN_ON(chunk->immutable);
 330
 331         pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
 332         if (!pages)
 333                 return -ENOMEM;
 334
 335         /* alloc and map */
 336         pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
 337                 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
 338                 if (rc)
 339                         goto err_free;
 340                 free_end = re;
 341         }
 342
 343         pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
 344                 rc = pcpu_map_pages(chunk, pages, populated, rs, re);
 345                 if (rc)
 346                         goto err_unmap;
 347                 unmap_end = re;
 348         }
 349         pcpu_post_map_flush(chunk, page_start, page_end);
 350
 351         /* commit new bitmap */
 352         bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
 353 clear:
 354         for_each_possible_cpu(cpu)
 355                 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
 356         return 0;
 357
 358 err_unmap:
 359         pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
 360         pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
 361                 pcpu_unmap_pages(chunk, pages, populated, rs, re);
 362         pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
 363 err_free:
 364         pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
 365                 pcpu_free_pages(chunk, pages, populated, rs, re);
 366         return rc;
 367 }
 368
 369 /**
 370  * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
 371  * @chunk: chunk to depopulate
 372  * @off: offset to the area to depopulate
 373  * @size: size of the area to depopulate in bytes
 374  * @flush: whether to flush cache and tlb or not
 375  *
 376  * For each cpu, depopulate and unmap pages [@page_start,@page_end)
 377  * from @chunk.  If @flush is true, vcache is flushed before unmapping
 378  * and tlb after.
 379  *
 380  * CONTEXT:
 381  * pcpu_alloc_mutex.
 382  */
 383 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
 384 {
 385         int page_start = PFN_DOWN(off);
 386         int page_end = PFN_UP(off + size);
 387         struct page **pages;
 388         unsigned long *populated;
 389         int rs, re;
 390
 391         /* quick path, check whether it's empty already */
 392         rs = page_start;
 393         pcpu_next_unpop(chunk, &rs, &re, page_end);
 394         if (rs == page_start && re == page_end)
 395                 return;
 396
 397         /* immutable chunks can't be depopulated */
 398         WARN_ON(chunk->immutable);
 399
 400         /*
 401          * If control reaches here, there must have been at least one
 402          * successful population attempt so the temp pages array must
 403          * be available now.
 404          */
 405         pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
 406         BUG_ON(!pages);
 407
 408         /* unmap and free */
 409         pcpu_pre_unmap_flush(chunk, page_start, page_end);
 410
 411         pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
 412                 pcpu_unmap_pages(chunk, pages, populated, rs, re);
 413
 414         /* no need to flush tlb, vmalloc will handle it lazily */
 415
 416         pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
 417                 pcpu_free_pages(chunk, pages, populated, rs, re);
 418
 419         /* commit new bitmap */
 420         bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
 421 }
 422
 423 static struct pcpu_chunk *pcpu_create_chunk(void)
 424 {
 425         struct pcpu_chunk *chunk;
 426         struct vm_struct **vms;
 427
 428         chunk = pcpu_alloc_chunk();
 429         if (!chunk)
 430                 return NULL;
 431
 432         vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
 433                                 pcpu_nr_groups, pcpu_atom_size);
 434         if (!vms) {
 435                 pcpu_free_chunk(chunk);
 436                 return NULL;
 437         }
 438
 439         chunk->data = vms;
 440         chunk->base_addr = vms[0]->addr - pcpu_group_offsets[0];
 441         return chunk;
 442 }
 443
 444 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk)
 445 {
 446         if (chunk && chunk->data)
 447                 pcpu_free_vm_areas(chunk->data, pcpu_nr_groups);
 448         pcpu_free_chunk(chunk);
 449 }
 450
 451 static struct page *pcpu_addr_to_page(void *addr)
 452 {
 453         return vmalloc_to_page(addr);
 454 }
 455
 456 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai)
 457 {
 458         /* no extra restriction */
 459         return 0;
 460 }