mm/percpu-vm.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * mm/percpu-vm.c - vmalloc area based chunk allocation
   4  *
   5  * Copyright (C) 2010           SUSE Linux Products GmbH
   6  * Copyright (C) 2010           Tejun Heo <tj@kernel.org>
   7  *
   8  * Chunks are mapped into vmalloc areas and populated page by page.
   9  * This is the default chunk allocator.
  10  */
  11
  12 static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
  13                                     unsigned int cpu, int page_idx)
  14 {
  15         /* must not be used on pre-mapped chunk */
  16         WARN_ON(chunk->immutable);
  17
  18         return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
  19 }
  20
  21 /**
  22  * pcpu_get_pages - get temp pages array
  23  *
  24  * Returns pointer to array of pointers to struct page which can be indexed
  25  * with pcpu_page_idx().  Note that there is only one array and accesses
  26  * should be serialized by pcpu_alloc_mutex.
  27  *
  28  * RETURNS:
  29  * Pointer to temp pages array on success.
  30  */
  31 static struct page **pcpu_get_pages(void)
  32 {
  33         static struct page **pages;
  34         size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
  35
  36         lockdep_assert_held(&pcpu_alloc_mutex);
  37
  38         if (!pages)
  39                 pages = pcpu_mem_zalloc(pages_size, GFP_KERNEL);
  40         return pages;
  41 }
  42
  43 /**
  44  * pcpu_free_pages - free pages which were allocated for @chunk
  45  * @chunk: chunk pages were allocated for
  46  * @pages: array of pages to be freed, indexed by pcpu_page_idx()
  47  * @page_start: page index of the first page to be freed
  48  * @page_end: page index of the last page to be freed + 1
  49  *
  50  * Free pages [@page_start and @page_end) in @pages for all units.
  51  * The pages were allocated for @chunk.
  52  */
  53 static void pcpu_free_pages(struct pcpu_chunk *chunk,
  54                             struct page **pages, int page_start, int page_end)
  55 {
  56         unsigned int cpu;
  57         int i;
  58
  59         for_each_possible_cpu(cpu) {
  60                 for (i = page_start; i < page_end; i++) {
  61                         struct page *page = pages[pcpu_page_idx(cpu, i)];
  62
  63                         if (page)
  64                                 __free_page(page);
  65                 }
  66         }
  67 }
  68
  69 /**
  70  * pcpu_alloc_pages - allocates pages for @chunk
  71  * @chunk: target chunk
  72  * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
  73  * @page_start: page index of the first page to be allocated
  74  * @page_end: page index of the last page to be allocated + 1
  75  * @gfp: allocation flags passed to the underlying allocator
  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                             gfp_t gfp)
  84 {
  85         unsigned int cpu, tcpu;
  86         int i;
  87
  88         gfp |= __GFP_HIGHMEM;
  89
  90         for_each_possible_cpu(cpu) {
  91                 for (i = page_start; i < page_end; i++) {
  92                         struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
  93
  94                         *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
  95                         if (!*pagep)
  96                                 goto err;
  97                 }
  98         }
  99         return 0;
 100
 101 err:
 102         while (--i >= page_start)
 103                 __free_page(pages[pcpu_page_idx(cpu, i)]);
 104
 105         for_each_possible_cpu(tcpu) {
 106                 if (tcpu == cpu)
 107                         break;
 108                 for (i = page_start; i < page_end; i++)
 109                         __free_page(pages[pcpu_page_idx(tcpu, i)]);
 110         }
 111         return -ENOMEM;
 112 }
 113
 114 /**
 115  * pcpu_pre_unmap_flush - flush cache prior to unmapping
 116  * @chunk: chunk the regions to be flushed belongs to
 117  * @page_start: page index of the first page to be flushed
 118  * @page_end: page index of the last page to be flushed + 1
 119  *
 120  * Pages in [@page_start,@page_end) of @chunk are about to be
 121  * unmapped.  Flush cache.  As each flushing trial can be very
 122  * expensive, issue flush on the whole region at once rather than
 123  * doing it for each cpu.  This could be an overkill but is more
 124  * scalable.
 125  */
 126 static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
 127                                  int page_start, int page_end)
 128 {
 129         flush_cache_vunmap(
 130                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 131                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 132 }
 133
 134 static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
 135 {
 136         unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
 137 }
 138
 139 /**
 140  * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
 141  * @chunk: chunk of interest
 142  * @pages: pages array which can be used to pass information to free
 143  * @page_start: page index of the first page to unmap
 144  * @page_end: page index of the last page to unmap + 1
 145  *
 146  * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
 147  * Corresponding elements in @pages were cleared by the caller and can
 148  * be used to carry information to pcpu_free_pages() which will be
 149  * called after all unmaps are finished.  The caller should call
 150  * proper pre/post flush functions.
 151  */
 152 static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
 153                              struct page **pages, int page_start, int page_end)
 154 {
 155         unsigned int cpu;
 156         int i;
 157
 158         for_each_possible_cpu(cpu) {
 159                 for (i = page_start; i < page_end; i++) {
 160                         struct page *page;
 161
 162                         page = pcpu_chunk_page(chunk, cpu, i);
 163                         WARN_ON(!page);
 164                         pages[pcpu_page_idx(cpu, i)] = page;
 165                 }
 166                 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
 167                                    page_end - page_start);
 168         }
 169 }
 170
 171 /**
 172  * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
 173  * @chunk: pcpu_chunk the regions to be flushed belong to
 174  * @page_start: page index of the first page to be flushed
 175  * @page_end: page index of the last page to be flushed + 1
 176  *
 177  * Pages [@page_start,@page_end) of @chunk have been unmapped.  Flush
 178  * TLB for the regions.  This can be skipped if the area is to be
 179  * returned to vmalloc as vmalloc will handle TLB flushing lazily.
 180  *
 181  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
 182  * for the whole region.
 183  */
 184 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
 185                                       int page_start, int page_end)
 186 {
 187         flush_tlb_kernel_range(
 188                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 189                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 190 }
 191
 192 static int __pcpu_map_pages(unsigned long addr, struct page **pages,
 193                             int nr_pages)
 194 {
 195         return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
 196                                         PAGE_KERNEL, pages);
 197 }
 198
 199 /**
 200  * pcpu_map_pages - map pages into a pcpu_chunk
 201  * @chunk: chunk of interest
 202  * @pages: pages array containing pages to be mapped
 203  * @page_start: page index of the first page to map
 204  * @page_end: page index of the last page to map + 1
 205  *
 206  * For each cpu, map pages [@page_start,@page_end) into @chunk.  The
 207  * caller is responsible for calling pcpu_post_map_flush() after all
 208  * mappings are complete.
 209  *
 210  * This function is responsible for setting up whatever is necessary for
 211  * reverse lookup (addr -> chunk).
 212  */
 213 static int pcpu_map_pages(struct pcpu_chunk *chunk,
 214                           struct page **pages, int page_start, int page_end)
 215 {
 216         unsigned int cpu, tcpu;
 217         int i, err;
 218
 219         for_each_possible_cpu(cpu) {
 220                 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
 221                                        &pages[pcpu_page_idx(cpu, page_start)],
 222                                        page_end - page_start);
 223                 if (err < 0)
 224                         goto err;
 225
 226                 for (i = page_start; i < page_end; i++)
 227                         pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
 228                                             chunk);
 229         }
 230         return 0;
 231 err:
 232         for_each_possible_cpu(tcpu) {
 233                 if (tcpu == cpu)
 234                         break;
 235                 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
 236                                    page_end - page_start);
 237         }
 238         pcpu_post_unmap_tlb_flush(chunk, page_start, page_end);
 239         return err;
 240 }
 241
 242 /**
 243  * pcpu_post_map_flush - flush cache after mapping
 244  * @chunk: pcpu_chunk the regions to be flushed belong to
 245  * @page_start: page index of the first page to be flushed
 246  * @page_end: page index of the last page to be flushed + 1
 247  *
 248  * Pages [@page_start,@page_end) of @chunk have been mapped.  Flush
 249  * cache.
 250  *
 251  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
 252  * for the whole region.
 253  */
 254 static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
 255                                 int page_start, int page_end)
 256 {
 257         flush_cache_vmap(
 258                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
 259                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
 260 }
 261
 262 /**
 263  * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
 264  * @chunk: chunk of interest
 265  * @page_start: the start page
 266  * @page_end: the end page
 267  * @gfp: allocation flags passed to the underlying memory allocator
 268  *
 269  * For each cpu, populate and map pages [@page_start,@page_end) into
 270  * @chunk.
 271  *
 272  * CONTEXT:
 273  * pcpu_alloc_mutex, does GFP_KERNEL allocation.
 274  */
 275 static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
 276                                int page_start, int page_end, gfp_t gfp)
 277 {
 278         struct page **pages;
 279
 280         pages = pcpu_get_pages();
 281         if (!pages)
 282                 return -ENOMEM;
 283
 284         if (pcpu_alloc_pages(chunk, pages, page_start, page_end, gfp))
 285                 return -ENOMEM;
 286
 287         if (pcpu_map_pages(chunk, pages, page_start, page_end)) {
 288                 pcpu_free_pages(chunk, pages, page_start, page_end);
 289                 return -ENOMEM;
 290         }
 291         pcpu_post_map_flush(chunk, page_start, page_end);
 292
 293         return 0;
 294 }
 295
 296 /**
 297  * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
 298  * @chunk: chunk to depopulate
 299  * @page_start: the start page
 300  * @page_end: the end page
 301  *
 302  * For each cpu, depopulate and unmap pages [@page_start,@page_end)
 303  * from @chunk.
 304  *
 305  * CONTEXT:
 306  * pcpu_alloc_mutex.
 307  */
 308 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
 309                                   int page_start, int page_end)
 310 {
 311         struct page **pages;
 312
 313         /*
 314          * If control reaches here, there must have been at least one
 315          * successful population attempt so the temp pages array must
 316          * be available now.
 317          */
 318         pages = pcpu_get_pages();
 319         BUG_ON(!pages);
 320
 321         /* unmap and free */
 322         pcpu_pre_unmap_flush(chunk, page_start, page_end);
 323
 324         pcpu_unmap_pages(chunk, pages, page_start, page_end);
 325
 326         /* no need to flush tlb, vmalloc will handle it lazily */
 327
 328         pcpu_free_pages(chunk, pages, page_start, page_end);
 329 }
 330
 331 static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp)
 332 {
 333         struct pcpu_chunk *chunk;
 334         struct vm_struct **vms;
 335
 336         chunk = pcpu_alloc_chunk(gfp);
 337         if (!chunk)
 338                 return NULL;
 339
 340         vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
 341                                 pcpu_nr_groups, pcpu_atom_size);
 342         if (!vms) {
 343                 pcpu_free_chunk(chunk);
 344                 return NULL;
 345         }
 346
 347         chunk->data = vms;
 348         chunk->base_addr = vms[0]->addr - pcpu_group_offsets[0];
 349
 350         pcpu_stats_chunk_alloc();
 351         trace_percpu_create_chunk(chunk->base_addr);
 352
 353         return chunk;
 354 }
 355
 356 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk)
 357 {
 358         if (!chunk)
 359                 return;
 360
 361         pcpu_stats_chunk_dealloc();
 362         trace_percpu_destroy_chunk(chunk->base_addr);
 363
 364         if (chunk->data)
 365                 pcpu_free_vm_areas(chunk->data, pcpu_nr_groups);
 366         pcpu_free_chunk(chunk);
 367 }
 368
 369 static struct page *pcpu_addr_to_page(void *addr)
 370 {
 371         return vmalloc_to_page(addr);
 372 }
 373
 374 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai)
 375 {
 376         /* no extra restriction */
 377         return 0;
 378 }