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         struct page *page = NULL;
 133
 134         mutex_lock(&mem_list_mutex);
 135
 136         list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list,
 137                         next) {
 138                 if ((mem->ua == ua) && (mem->entries == entries)) {
 139                         ++mem->used;
 140                         *pmem = mem;
 141                         goto unlock_exit;
 142                 }
 143
 144                 /* Overlap? */
 145                 if ((mem->ua < (ua + (entries << PAGE_SHIFT))) &&
 146                                 (ua < (mem->ua +
 147                                        (mem->entries << PAGE_SHIFT)))) {
 148                         ret = -EINVAL;
 149                         goto unlock_exit;
 150                 }
 151
 152         }
 153
 154         ret = mm_iommu_adjust_locked_vm(mm, entries, true);
 155         if (ret)
 156                 goto unlock_exit;
 157
 158         locked_entries = entries;
 159
 160         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 161         if (!mem) {
 162                 ret = -ENOMEM;
 163                 goto unlock_exit;
 164         }
 165
 166         /*
 167          * For a starting point for a maximum page size calculation
 168          * we use @ua and @entries natural alignment to allow IOMMU pages
 169          * smaller than huge pages but still bigger than PAGE_SIZE.
 170          */
 171         mem->pageshift = __ffs(ua | (entries << PAGE_SHIFT));
 172         mem->hpas = vzalloc(array_size(entries, sizeof(mem->hpas[0])));
 173         if (!mem->hpas) {
 174                 kfree(mem);
 175                 ret = -ENOMEM;
 176                 goto unlock_exit;
 177         }
 178
 179         for (i = 0; i < entries; ++i) {
 180                 if (1 != get_user_pages_fast(ua + (i << PAGE_SHIFT),
 181                                         1/* pages */, 1/* iswrite */, &page)) {
 182                         ret = -EFAULT;
 183                         for (j = 0; j < i; ++j)
 184                                 put_page(pfn_to_page(mem->hpas[j] >>
 185                                                 PAGE_SHIFT));
 186                         vfree(mem->hpas);
 187                         kfree(mem);
 188                         goto unlock_exit;
 189                 }
 190                 /*
 191                  * If we get a page from the CMA zone, since we are going to
 192                  * be pinning these entries, we might as well move them out
 193                  * of the CMA zone if possible. NOTE: faulting in + migration
 194                  * can be expensive. Batching can be considered later
 195                  */
 196                 if (is_migrate_cma_page(page)) {
 197                         if (mm_iommu_move_page_from_cma(page))
 198                                 goto populate;
 199                         if (1 != get_user_pages_fast(ua + (i << PAGE_SHIFT),
 200                                                 1/* pages */, 1/* iswrite */,
 201                                                 &page)) {
 202                                 ret = -EFAULT;
 203                                 for (j = 0; j < i; ++j)
 204                                         put_page(pfn_to_page(mem->hpas[j] >>
 205                                                                 PAGE_SHIFT));
 206                                 vfree(mem->hpas);
 207                                 kfree(mem);
 208                                 goto unlock_exit;
 209                         }
 210                 }
 211 populate:
 212                 pageshift = PAGE_SHIFT;
 213                 if (PageCompound(page)) {
 214                         pte_t *pte;
 215                         struct page *head = compound_head(page);
 216                         unsigned int compshift = compound_order(head);
 217
 218                         local_irq_save(flags); /* disables as well */
 219                         pte = find_linux_pte(mm->pgd, ua, NULL, &pageshift);
 220                         local_irq_restore(flags);
 221
 222                         /* Double check it is still the same pinned page */
 223                         if (pte && pte_page(*pte) == head &&
 224                                         pageshift == compshift)
 225                                 pageshift = max_t(unsigned int, pageshift,
 226                                                 PAGE_SHIFT);
 227                 }
 228                 mem->pageshift = min(mem->pageshift, pageshift);
 229                 mem->hpas[i] = page_to_pfn(page) << PAGE_SHIFT;
 230         }
 231
 232         atomic64_set(&mem->mapped, 1);
 233         mem->used = 1;
 234         mem->ua = ua;
 235         mem->entries = entries;
 236         *pmem = mem;
 237
 238         list_add_rcu(&mem->next, &mm->context.iommu_group_mem_list);
 239
 240 unlock_exit:
 241         if (locked_entries && ret)
 242                 mm_iommu_adjust_locked_vm(mm, locked_entries, false);
 243
 244         mutex_unlock(&mem_list_mutex);
 245
 246         return ret;
 247 }
 248 EXPORT_SYMBOL_GPL(mm_iommu_get);
 249
 250 static void mm_iommu_unpin(struct mm_iommu_table_group_mem_t *mem)
 251 {
 252         long i;
 253         struct page *page = NULL;
 254
 255         for (i = 0; i < mem->entries; ++i) {
 256                 if (!mem->hpas[i])
 257                         continue;
 258
 259                 page = pfn_to_page(mem->hpas[i] >> PAGE_SHIFT);
 260                 if (!page)
 261                         continue;
 262
 263                 put_page(page);
 264                 mem->hpas[i] = 0;
 265         }
 266 }
 267
 268 static void mm_iommu_do_free(struct mm_iommu_table_group_mem_t *mem)
 269 {
 270
 271         mm_iommu_unpin(mem);
 272         vfree(mem->hpas);
 273         kfree(mem);
 274 }
 275
 276 static void mm_iommu_free(struct rcu_head *head)
 277 {
 278         struct mm_iommu_table_group_mem_t *mem = container_of(head,
 279                         struct mm_iommu_table_group_mem_t, rcu);
 280
 281         mm_iommu_do_free(mem);
 282 }
 283
 284 static void mm_iommu_release(struct mm_iommu_table_group_mem_t *mem)
 285 {
 286         list_del_rcu(&mem->next);
 287         call_rcu(&mem->rcu, mm_iommu_free);
 288 }
 289
 290 long mm_iommu_put(struct mm_struct *mm, struct mm_iommu_table_group_mem_t *mem)
 291 {
 292         long ret = 0;
 293
 294         mutex_lock(&mem_list_mutex);
 295
 296         if (mem->used == 0) {
 297                 ret = -ENOENT;
 298                 goto unlock_exit;
 299         }
 300
 301         --mem->used;
 302         /* There are still users, exit */
 303         if (mem->used)
 304                 goto unlock_exit;
 305
 306         /* Are there still mappings? */
 307         if (atomic_cmpxchg(&mem->mapped, 1, 0) != 1) {
 308                 ++mem->used;
 309                 ret = -EBUSY;
 310                 goto unlock_exit;
 311         }
 312
 313         /* @mapped became 0 so now mappings are disabled, release the region */
 314         mm_iommu_release(mem);
 315
 316         mm_iommu_adjust_locked_vm(mm, mem->entries, false);
 317
 318 unlock_exit:
 319         mutex_unlock(&mem_list_mutex);
 320
 321         return ret;
 322 }
 323 EXPORT_SYMBOL_GPL(mm_iommu_put);
 324
 325 struct mm_iommu_table_group_mem_t *mm_iommu_lookup(struct mm_struct *mm,
 326                 unsigned long ua, unsigned long size)
 327 {
 328         struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
 329
 330         list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list, next) {
 331                 if ((mem->ua <= ua) &&
 332                                 (ua + size <= mem->ua +
 333                                  (mem->entries << PAGE_SHIFT))) {
 334                         ret = mem;
 335                         break;
 336                 }
 337         }
 338
 339         return ret;
 340 }
 341 EXPORT_SYMBOL_GPL(mm_iommu_lookup);
 342
 343 struct mm_iommu_table_group_mem_t *mm_iommu_lookup_rm(struct mm_struct *mm,
 344                 unsigned long ua, unsigned long size)
 345 {
 346         struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
 347
 348         list_for_each_entry_lockless(mem, &mm->context.iommu_group_mem_list,
 349                         next) {
 350                 if ((mem->ua <= ua) &&
 351                                 (ua + size <= mem->ua +
 352                                  (mem->entries << PAGE_SHIFT))) {
 353                         ret = mem;
 354                         break;
 355                 }
 356         }
 357
 358         return ret;
 359 }
 360 EXPORT_SYMBOL_GPL(mm_iommu_lookup_rm);
 361
 362 struct mm_iommu_table_group_mem_t *mm_iommu_find(struct mm_struct *mm,
 363                 unsigned long ua, unsigned long entries)
 364 {
 365         struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
 366
 367         list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list, next) {
 368                 if ((mem->ua == ua) && (mem->entries == entries)) {
 369                         ret = mem;
 370                         break;
 371                 }
 372         }
 373
 374         return ret;
 375 }
 376 EXPORT_SYMBOL_GPL(mm_iommu_find);
 377
 378 long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem,
 379                 unsigned long ua, unsigned int pageshift, unsigned long *hpa)
 380 {
 381         const long entry = (ua - mem->ua) >> PAGE_SHIFT;
 382         u64 *va = &mem->hpas[entry];
 383
 384         if (entry >= mem->entries)
 385                 return -EFAULT;
 386
 387         if (pageshift > mem->pageshift)
 388                 return -EFAULT;
 389
 390         *hpa = *va | (ua & ~PAGE_MASK);
 391
 392         return 0;
 393 }
 394 EXPORT_SYMBOL_GPL(mm_iommu_ua_to_hpa);
 395
 396 long mm_iommu_ua_to_hpa_rm(struct mm_iommu_table_group_mem_t *mem,
 397                 unsigned long ua, unsigned int pageshift, unsigned long *hpa)
 398 {
 399         const long entry = (ua - mem->ua) >> PAGE_SHIFT;
 400         void *va = &mem->hpas[entry];
 401         unsigned long *pa;
 402
 403         if (entry >= mem->entries)
 404                 return -EFAULT;
 405
 406         if (pageshift > mem->pageshift)
 407                 return -EFAULT;
 408
 409         pa = (void *) vmalloc_to_phys(va);
 410         if (!pa)
 411                 return -EFAULT;
 412
 413         *hpa = *pa | (ua & ~PAGE_MASK);
 414
 415         return 0;
 416 }
 417 EXPORT_SYMBOL_GPL(mm_iommu_ua_to_hpa_rm);
 418
 419 long mm_iommu_mapped_inc(struct mm_iommu_table_group_mem_t *mem)
 420 {
 421         if (atomic64_inc_not_zero(&mem->mapped))
 422                 return 0;
 423
 424         /* Last mm_iommu_put() has been called, no more mappings allowed() */
 425         return -ENXIO;
 426 }
 427 EXPORT_SYMBOL_GPL(mm_iommu_mapped_inc);
 428
 429 void mm_iommu_mapped_dec(struct mm_iommu_table_group_mem_t *mem)
 430 {
 431         atomic64_add_unless(&mem->mapped, -1, 1);
 432 }
 433 EXPORT_SYMBOL_GPL(mm_iommu_mapped_dec);
 434
 435 void mm_iommu_init(struct mm_struct *mm)
 436 {
 437         INIT_LIST_HEAD_RCU(&mm->context.iommu_group_mem_list);
 438 }