arch/powerpc/mm/mmu_context_iommu.c

   1 /*
   2  *  IOMMU helpers in MMU context.
   3  *
   4  *  Copyright (C) 2015 IBM Corp. <aik@ozlabs.ru>
   5  *
   6  *  This program is free software; you can redistribute it and/or
   7  *  modify it under the terms of the GNU General Public License
   8  *  as published by the Free Software Foundation; either version
   9  *  2 of the License, or (at your option) any later version.
  10  *
  11  */
  12
  13 #include <linux/sched/signal.h>
  14 #include <linux/slab.h>
  15 #include <linux/rculist.h>
  16 #include <linux/vmalloc.h>
  17 #include <linux/mutex.h>
  18 #include <linux/migrate.h>
  19 #include <linux/hugetlb.h>
  20 #include <linux/swap.h>
  21 #include <asm/mmu_context.h>
  22 #include <asm/pte-walk.h>
  23
  24 static DEFINE_MUTEX(mem_list_mutex);
  25
  26 struct mm_iommu_table_group_mem_t {
  27         struct list_head next;
  28         struct rcu_head rcu;
  29         unsigned long used;
  30         atomic64_t mapped;
  31         unsigned int pageshift;
  32         u64 ua;                 /* userspace address */
  33         u64 entries;            /* number of entries in hpas[] */
  34         u64 *hpas;              /* vmalloc'ed */
  35 };
  36
  37 static long mm_iommu_adjust_locked_vm(struct mm_struct *mm,
  38                 unsigned long npages, bool incr)
  39 {
  40         long ret = 0, locked, lock_limit;
  41
  42         if (!npages)
  43                 return 0;
  44
  45         down_write(&mm->mmap_sem);
  46
  47         if (incr) {
  48                 locked = mm->locked_vm + npages;
  49                 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
  50                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
  51                         ret = -ENOMEM;
  52                 else
  53                         mm->locked_vm += npages;
  54         } else {
  55                 if (WARN_ON_ONCE(npages > mm->locked_vm))
  56                         npages = mm->locked_vm;
  57                 mm->locked_vm -= npages;
  58         }
  59
  60         pr_debug("[%d] RLIMIT_MEMLOCK HASH64 %c%ld %ld/%ld\n",
  61                         current ? current->pid : 0,
  62                         incr ? '+' : '-',
  63                         npages << PAGE_SHIFT,
  64                         mm->locked_vm << PAGE_SHIFT,
  65                         rlimit(RLIMIT_MEMLOCK));
  66         up_write(&mm->mmap_sem);
  67
  68         return ret;
  69 }
  70
  71 bool mm_iommu_preregistered(struct mm_struct *mm)
  72 {
  73         return !list_empty(&mm->context.iommu_group_mem_list);
  74 }
  75 EXPORT_SYMBOL_GPL(mm_iommu_preregistered);
  76
  77 /*
  78  * Taken from alloc_migrate_target with changes to remove CMA allocations
  79  */
  80 struct page *new_iommu_non_cma_page(struct page *page, unsigned long private)
  81 {
  82         gfp_t gfp_mask = GFP_USER;
  83         struct page *new_page;
  84
  85         if (PageCompound(page))
  86                 return NULL;
  87
  88         if (PageHighMem(page))
  89                 gfp_mask |= __GFP_HIGHMEM;
  90
  91         /*
  92          * We don't want the allocation to force an OOM if possibe
  93          */
  94         new_page = alloc_page(gfp_mask | __GFP_NORETRY | __GFP_NOWARN);
  95         return new_page;
  96 }
  97
  98 static int mm_iommu_move_page_from_cma(struct page *page)
  99 {
 100         int ret = 0;
 101         LIST_HEAD(cma_migrate_pages);
 102
 103         /* Ignore huge pages for now */
 104         if (PageCompound(page))
 105                 return -EBUSY;
 106
 107         lru_add_drain();
 108         ret = isolate_lru_page(page);
 109         if (ret)
 110                 return ret;
 111
 112         list_add(&page->lru, &cma_migrate_pages);
 113         put_page(page); /* Drop the gup reference */
 114
 115         ret = migrate_pages(&cma_migrate_pages, new_iommu_non_cma_page,
 116                                 NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE);
 117         if (ret) {
 118                 if (!list_empty(&cma_migrate_pages))
 119                         putback_movable_pages(&cma_migrate_pages);
 120         }
 121
 122         return 0;
 123 }
 124
 125 long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries,
 126                 struct mm_iommu_table_group_mem_t **pmem)
 127 {
 128         struct mm_iommu_table_group_mem_t *mem;
 129         long i, j, ret = 0, locked_entries = 0;
 130         unsigned int pageshift;
 131         unsigned long flags;
 132         unsigned long cur_ua;
 133         struct page *page = NULL;
 134
 135         mutex_lock(&mem_list_mutex);
 136
 137         list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list,
 138                         next) {
 139                 if ((mem->ua == ua) && (mem->entries == entries)) {
 140                         ++mem->used;
 141                         *pmem = mem;
 142                         goto unlock_exit;
 143                 }
 144
 145                 /* Overlap? */
 146                 if ((mem->ua < (ua + (entries << PAGE_SHIFT))) &&
 147                                 (ua < (mem->ua +
 148                                        (mem->entries << PAGE_SHIFT)))) {
 149                         ret = -EINVAL;
 150                         goto unlock_exit;
 151                 }
 152
 153         }
 154
 155         ret = mm_iommu_adjust_locked_vm(mm, entries, true);
 156         if (ret)
 157                 goto unlock_exit;
 158
 159         locked_entries = entries;
 160
 161         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 162         if (!mem) {
 163                 ret = -ENOMEM;
 164                 goto unlock_exit;
 165         }
 166
 167         /*
 168          * For a starting point for a maximum page size calculation
 169          * we use @ua and @entries natural alignment to allow IOMMU pages
 170          * smaller than huge pages but still bigger than PAGE_SIZE.
 171          */
 172         mem->pageshift = __ffs(ua | (entries << PAGE_SHIFT));
 173         mem->hpas = vzalloc(array_size(entries, sizeof(mem->hpas[0])));
 174         if (!mem->hpas) {
 175                 kfree(mem);
 176                 ret = -ENOMEM;
 177                 goto unlock_exit;
 178         }
 179
 180         for (i = 0; i < entries; ++i) {
 181                 cur_ua = ua + (i << PAGE_SHIFT);
 182                 if (1 != get_user_pages_fast(cur_ua,
 183                                         1/* pages */, 1/* iswrite */, &page)) {
 184                         ret = -EFAULT;
 185                         for (j = 0; j < i; ++j)
 186                                 put_page(pfn_to_page(mem->hpas[j] >>
 187                                                 PAGE_SHIFT));
 188                         vfree(mem->hpas);
 189                         kfree(mem);
 190                         goto unlock_exit;
 191                 }
 192                 /*
 193                  * If we get a page from the CMA zone, since we are going to
 194                  * be pinning these entries, we might as well move them out
 195                  * of the CMA zone if possible. NOTE: faulting in + migration
 196                  * can be expensive. Batching can be considered later
 197                  */
 198                 if (is_migrate_cma_page(page)) {
 199                         if (mm_iommu_move_page_from_cma(page))
 200                                 goto populate;
 201                         if (1 != get_user_pages_fast(cur_ua,
 202                                                 1/* pages */, 1/* iswrite */,
 203                                                 &page)) {
 204                                 ret = -EFAULT;
 205                                 for (j = 0; j < i; ++j)
 206                                         put_page(pfn_to_page(mem->hpas[j] >>
 207                                                                 PAGE_SHIFT));
 208                                 vfree(mem->hpas);
 209                                 kfree(mem);
 210                                 goto unlock_exit;
 211                         }
 212                 }
 213 populate:
 214                 pageshift = PAGE_SHIFT;
 215                 if (mem->pageshift > PAGE_SHIFT && PageCompound(page)) {
 216                         pte_t *pte;
 217                         struct page *head = compound_head(page);
 218                         unsigned int compshift = compound_order(head);
 219                         unsigned int pteshift;
 220
 221                         local_irq_save(flags); /* disables as well */
 222                         pte = find_linux_pte(mm->pgd, cur_ua, NULL, &pteshift);
 223
 224                         /* Double check it is still the same pinned page */
 225                         if (pte && pte_page(*pte) == head &&
 226                             pteshift == compshift + PAGE_SHIFT)
 227                                 pageshift = max_t(unsigned int, pteshift,
 228                                                 PAGE_SHIFT);
 229                         local_irq_restore(flags);
 230                 }
 231                 mem->pageshift = min(mem->pageshift, pageshift);
 232                 mem->hpas[i] = page_to_pfn(page) << PAGE_SHIFT;
 233         }
 234
 235         atomic64_set(&mem->mapped, 1);
 236         mem->used = 1;
 237         mem->ua = ua;
 238         mem->entries = entries;
 239         *pmem = mem;
 240
 241         list_add_rcu(&mem->next, &mm->context.iommu_group_mem_list);
 242
 243 unlock_exit:
 244         if (locked_entries && ret)
 245                 mm_iommu_adjust_locked_vm(mm, locked_entries, false);
 246
 247         mutex_unlock(&mem_list_mutex);
 248
 249         return ret;
 250 }
 251 EXPORT_SYMBOL_GPL(mm_iommu_get);
 252
 253 static void mm_iommu_unpin(struct mm_iommu_table_group_mem_t *mem)
 254 {
 255         long i;
 256         struct page *page = NULL;
 257
 258         for (i = 0; i < mem->entries; ++i) {
 259                 if (!mem->hpas[i])
 260                         continue;
 261
 262                 page = pfn_to_page(mem->hpas[i] >> PAGE_SHIFT);
 263                 if (!page)
 264                         continue;
 265
 266                 put_page(page);
 267                 mem->hpas[i] = 0;
 268         }
 269 }
 270
 271 static void mm_iommu_do_free(struct mm_iommu_table_group_mem_t *mem)
 272 {
 273
 274         mm_iommu_unpin(mem);
 275         vfree(mem->hpas);
 276         kfree(mem);
 277 }
 278
 279 static void mm_iommu_free(struct rcu_head *head)
 280 {
 281         struct mm_iommu_table_group_mem_t *mem = container_of(head,
 282                         struct mm_iommu_table_group_mem_t, rcu);
 283
 284         mm_iommu_do_free(mem);
 285 }
 286
 287 static void mm_iommu_release(struct mm_iommu_table_group_mem_t *mem)
 288 {
 289         list_del_rcu(&mem->next);
 290         call_rcu(&mem->rcu, mm_iommu_free);
 291 }
 292
 293 long mm_iommu_put(struct mm_struct *mm, struct mm_iommu_table_group_mem_t *mem)
 294 {
 295         long ret = 0;
 296
 297         mutex_lock(&mem_list_mutex);
 298
 299         if (mem->used == 0) {
 300                 ret = -ENOENT;
 301                 goto unlock_exit;
 302         }
 303
 304         --mem->used;
 305         /* There are still users, exit */
 306         if (mem->used)
 307                 goto unlock_exit;
 308
 309         /* Are there still mappings? */
 310         if (atomic_cmpxchg(&mem->mapped, 1, 0) != 1) {
 311                 ++mem->used;
 312                 ret = -EBUSY;
 313                 goto unlock_exit;
 314         }
 315
 316         /* @mapped became 0 so now mappings are disabled, release the region */
 317         mm_iommu_release(mem);
 318
 319         mm_iommu_adjust_locked_vm(mm, mem->entries, false);
 320
 321 unlock_exit:
 322         mutex_unlock(&mem_list_mutex);
 323
 324         return ret;
 325 }
 326 EXPORT_SYMBOL_GPL(mm_iommu_put);
 327
 328 struct mm_iommu_table_group_mem_t *mm_iommu_lookup(struct mm_struct *mm,
 329                 unsigned long ua, unsigned long size)
 330 {
 331         struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
 332
 333         list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list, next) {
 334                 if ((mem->ua <= ua) &&
 335                                 (ua + size <= mem->ua +
 336                                  (mem->entries << PAGE_SHIFT))) {
 337                         ret = mem;
 338                         break;
 339                 }
 340         }
 341
 342         return ret;
 343 }
 344 EXPORT_SYMBOL_GPL(mm_iommu_lookup);
 345
 346 struct mm_iommu_table_group_mem_t *mm_iommu_lookup_rm(struct mm_struct *mm,
 347                 unsigned long ua, unsigned long size)
 348 {
 349         struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
 350
 351         list_for_each_entry_lockless(mem, &mm->context.iommu_group_mem_list,
 352                         next) {
 353                 if ((mem->ua <= ua) &&
 354                                 (ua + size <= mem->ua +
 355                                  (mem->entries << PAGE_SHIFT))) {
 356                         ret = mem;
 357                         break;
 358                 }
 359         }
 360
 361         return ret;
 362 }
 363 EXPORT_SYMBOL_GPL(mm_iommu_lookup_rm);
 364
 365 struct mm_iommu_table_group_mem_t *mm_iommu_find(struct mm_struct *mm,
 366                 unsigned long ua, unsigned long entries)
 367 {
 368         struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
 369
 370         list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list, next) {
 371                 if ((mem->ua == ua) && (mem->entries == entries)) {
 372                         ret = mem;
 373                         break;
 374                 }
 375         }
 376
 377         return ret;
 378 }
 379 EXPORT_SYMBOL_GPL(mm_iommu_find);
 380
 381 long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem,
 382                 unsigned long ua, unsigned int pageshift, unsigned long *hpa)
 383 {
 384         const long entry = (ua - mem->ua) >> PAGE_SHIFT;
 385         u64 *va = &mem->hpas[entry];
 386
 387         if (entry >= mem->entries)
 388                 return -EFAULT;
 389
 390         if (pageshift > mem->pageshift)
 391                 return -EFAULT;
 392
 393         *hpa = *va | (ua & ~PAGE_MASK);
 394
 395         return 0;
 396 }
 397 EXPORT_SYMBOL_GPL(mm_iommu_ua_to_hpa);
 398
 399 long mm_iommu_ua_to_hpa_rm(struct mm_iommu_table_group_mem_t *mem,
 400                 unsigned long ua, unsigned int pageshift, unsigned long *hpa)
 401 {
 402         const long entry = (ua - mem->ua) >> PAGE_SHIFT;
 403         void *va = &mem->hpas[entry];
 404         unsigned long *pa;
 405
 406         if (entry >= mem->entries)
 407                 return -EFAULT;
 408
 409         if (pageshift > mem->pageshift)
 410                 return -EFAULT;
 411
 412         pa = (void *) vmalloc_to_phys(va);
 413         if (!pa)
 414                 return -EFAULT;
 415
 416         *hpa = *pa | (ua & ~PAGE_MASK);
 417
 418         return 0;
 419 }
 420 EXPORT_SYMBOL_GPL(mm_iommu_ua_to_hpa_rm);
 421
 422 long mm_iommu_mapped_inc(struct mm_iommu_table_group_mem_t *mem)
 423 {
 424         if (atomic64_inc_not_zero(&mem->mapped))
 425                 return 0;
 426
 427         /* Last mm_iommu_put() has been called, no more mappings allowed() */
 428         return -ENXIO;
 429 }
 430 EXPORT_SYMBOL_GPL(mm_iommu_mapped_inc);
 431
 432 void mm_iommu_mapped_dec(struct mm_iommu_table_group_mem_t *mem)
 433 {
 434         atomic64_add_unless(&mem->mapped, -1, 1);
 435 }
 436 EXPORT_SYMBOL_GPL(mm_iommu_mapped_dec);
 437
 438 void mm_iommu_init(struct mm_struct *mm)
 439 {
 440         INIT_LIST_HEAD_RCU(&mm->context.iommu_group_mem_list);
 441 }