mm/userfaultfd.c

   1 /*
   2  *  mm/userfaultfd.c
   3  *
   4  *  Copyright (C) 2015  Red Hat, Inc.
   5  *
   6  *  This work is licensed under the terms of the GNU GPL, version 2. See
   7  *  the COPYING file in the top-level directory.
   8  */
   9
  10 #include <linux/mm.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/rmap.h>
  13 #include <linux/swap.h>
  14 #include <linux/swapops.h>
  15 #include <linux/userfaultfd_k.h>
  16 #include <linux/mmu_notifier.h>
  17 #include <linux/hugetlb.h>
  18 #include <linux/pagemap.h>
  19 #include <asm/tlbflush.h>
  20 #include "internal.h"
  21
  22 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
  23                             pmd_t *dst_pmd,
  24                             struct vm_area_struct *dst_vma,
  25                             unsigned long dst_addr,
  26                             unsigned long src_addr,
  27                             struct page **pagep)
  28 {
  29         struct mem_cgroup *memcg;
  30         pte_t _dst_pte, *dst_pte;
  31         spinlock_t *ptl;
  32         void *page_kaddr;
  33         int ret;
  34         struct page *page;
  35
  36         if (!*pagep) {
  37                 ret = -ENOMEM;
  38                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
  39                 if (!page)
  40                         goto out;
  41
  42                 page_kaddr = kmap_atomic(page);
  43                 ret = copy_from_user(page_kaddr,
  44                                      (const void __user *) src_addr,
  45                                      PAGE_SIZE);
  46                 kunmap_atomic(page_kaddr);
  47
  48                 /* fallback to copy_from_user outside mmap_sem */
  49                 if (unlikely(ret)) {
  50                         ret = -EFAULT;
  51                         *pagep = page;
  52                         /* don't free the page */
  53                         goto out;
  54                 }
  55         } else {
  56                 page = *pagep;
  57                 *pagep = NULL;
  58         }
  59
  60         /*
  61          * The memory barrier inside __SetPageUptodate makes sure that
  62          * preceeding stores to the page contents become visible before
  63          * the set_pte_at() write.
  64          */
  65         __SetPageUptodate(page);
  66
  67         ret = -ENOMEM;
  68         if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
  69                 goto out_release;
  70
  71         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  72         if (dst_vma->vm_flags & VM_WRITE)
  73                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
  74
  75         ret = -EEXIST;
  76         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  77         if (!pte_none(*dst_pte))
  78                 goto out_release_uncharge_unlock;
  79
  80         inc_mm_counter(dst_mm, MM_ANONPAGES);
  81         page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
  82         mem_cgroup_commit_charge(page, memcg, false, false);
  83         lru_cache_add_active_or_unevictable(page, dst_vma);
  84
  85         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
  86
  87         /* No need to invalidate - it was non-present before */
  88         update_mmu_cache(dst_vma, dst_addr, dst_pte);
  89
  90         pte_unmap_unlock(dst_pte, ptl);
  91         ret = 0;
  92 out:
  93         return ret;
  94 out_release_uncharge_unlock:
  95         pte_unmap_unlock(dst_pte, ptl);
  96         mem_cgroup_cancel_charge(page, memcg, false);
  97 out_release:
  98         put_page(page);
  99         goto out;
 100 }
 101
 102 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 103                               pmd_t *dst_pmd,
 104                               struct vm_area_struct *dst_vma,
 105                               unsigned long dst_addr)
 106 {
 107         pte_t _dst_pte, *dst_pte;
 108         spinlock_t *ptl;
 109         int ret;
 110
 111         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 112                                          dst_vma->vm_page_prot));
 113         ret = -EEXIST;
 114         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 115         if (!pte_none(*dst_pte))
 116                 goto out_unlock;
 117         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 118         /* No need to invalidate - it was non-present before */
 119         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 120         ret = 0;
 121 out_unlock:
 122         pte_unmap_unlock(dst_pte, ptl);
 123         return ret;
 124 }
 125
 126 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 127 {
 128         pgd_t *pgd;
 129         pud_t *pud;
 130         pmd_t *pmd = NULL;
 131
 132         pgd = pgd_offset(mm, address);
 133         pud = pud_alloc(mm, pgd, address);
 134         if (pud)
 135                 /*
 136                  * Note that we didn't run this because the pmd was
 137                  * missing, the *pmd may be already established and in
 138                  * turn it may also be a trans_huge_pmd.
 139                  */
 140                 pmd = pmd_alloc(mm, pud, address);
 141         return pmd;
 142 }
 143
 144 #ifdef CONFIG_HUGETLB_PAGE
 145 /*
 146  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 147  * called with mmap_sem held, it will release mmap_sem before returning.
 148  */
 149 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 150                                               struct vm_area_struct *dst_vma,
 151                                               unsigned long dst_start,
 152                                               unsigned long src_start,
 153                                               unsigned long len,
 154                                               bool zeropage)
 155 {
 156         ssize_t err;
 157         pte_t *dst_pte;
 158         unsigned long src_addr, dst_addr;
 159         long copied;
 160         struct page *page;
 161         struct hstate *h;
 162         unsigned long vma_hpagesize;
 163         pgoff_t idx;
 164         u32 hash;
 165         struct address_space *mapping;
 166
 167         /*
 168          * There is no default zero huge page for all huge page sizes as
 169          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 170          * by THP.  Since we can not reliably insert a zero page, this
 171          * feature is not supported.
 172          */
 173         if (zeropage) {
 174                 up_read(&dst_mm->mmap_sem);
 175                 return -EINVAL;
 176         }
 177
 178         src_addr = src_start;
 179         dst_addr = dst_start;
 180         copied = 0;
 181         page = NULL;
 182         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 183
 184         /*
 185          * Validate alignment based on huge page size
 186          */
 187         err = -EINVAL;
 188         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 189                 goto out_unlock;
 190
 191 retry:
 192         /*
 193          * On routine entry dst_vma is set.  If we had to drop mmap_sem and
 194          * retry, dst_vma will be set to NULL and we must lookup again.
 195          */
 196         if (!dst_vma) {
 197                 err = -EINVAL;
 198                 dst_vma = find_vma(dst_mm, dst_start);
 199                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 200                         goto out_unlock;
 201
 202                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 203                         goto out_unlock;
 204
 205                 /*
 206                  * Make sure the vma is not shared, that the remaining dst
 207                  * range is both valid and fully within a single existing vma.
 208                  */
 209                 if (dst_vma->vm_flags & VM_SHARED)
 210                         goto out_unlock;
 211                 if (dst_start < dst_vma->vm_start ||
 212                     dst_start + len > dst_vma->vm_end)
 213                         goto out_unlock;
 214         }
 215
 216         if (WARN_ON(dst_addr & (vma_hpagesize - 1) ||
 217                     (len - copied) & (vma_hpagesize - 1)))
 218                 goto out_unlock;
 219
 220         /*
 221          * Only allow __mcopy_atomic_hugetlb on userfaultfd registered ranges.
 222          */
 223         if (!dst_vma->vm_userfaultfd_ctx.ctx)
 224                 goto out_unlock;
 225
 226         /*
 227          * Ensure the dst_vma has a anon_vma.
 228          */
 229         err = -ENOMEM;
 230         if (unlikely(anon_vma_prepare(dst_vma)))
 231                 goto out_unlock;
 232
 233         h = hstate_vma(dst_vma);
 234
 235         while (src_addr < src_start + len) {
 236                 pte_t dst_pteval;
 237
 238                 BUG_ON(dst_addr >= dst_start + len);
 239                 VM_BUG_ON(dst_addr & ~huge_page_mask(h));
 240
 241                 /*
 242                  * Serialize via hugetlb_fault_mutex
 243                  */
 244                 idx = linear_page_index(dst_vma, dst_addr);
 245                 mapping = dst_vma->vm_file->f_mapping;
 246                 hash = hugetlb_fault_mutex_hash(h, dst_mm, dst_vma, mapping,
 247                                                                 idx, dst_addr);
 248                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 249
 250                 err = -ENOMEM;
 251                 dst_pte = huge_pte_alloc(dst_mm, dst_addr, huge_page_size(h));
 252                 if (!dst_pte) {
 253                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 254                         goto out_unlock;
 255                 }
 256
 257                 err = -EEXIST;
 258                 dst_pteval = huge_ptep_get(dst_pte);
 259                 if (!huge_pte_none(dst_pteval)) {
 260                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 261                         goto out_unlock;
 262                 }
 263
 264                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 265                                                 dst_addr, src_addr, &page);
 266
 267                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 268
 269                 cond_resched();
 270
 271                 if (unlikely(err == -EFAULT)) {
 272                         up_read(&dst_mm->mmap_sem);
 273                         BUG_ON(!page);
 274
 275                         err = copy_huge_page_from_user(page,
 276                                                 (const void __user *)src_addr,
 277                                                 pages_per_huge_page(h), true);
 278                         if (unlikely(err)) {
 279                                 err = -EFAULT;
 280                                 goto out;
 281                         }
 282                         down_read(&dst_mm->mmap_sem);
 283
 284                         dst_vma = NULL;
 285                         goto retry;
 286                 } else
 287                         BUG_ON(page);
 288
 289                 if (!err) {
 290                         dst_addr += vma_hpagesize;
 291                         src_addr += vma_hpagesize;
 292                         copied += vma_hpagesize;
 293
 294                         if (fatal_signal_pending(current))
 295                                 err = -EINTR;
 296                 }
 297                 if (err)
 298                         break;
 299         }
 300
 301 out_unlock:
 302         up_read(&dst_mm->mmap_sem);
 303 out:
 304         if (page) {
 305                 /*
 306                  * We encountered an error and are about to free a newly
 307                  * allocated huge page.  It is possible that there was a
 308                  * reservation associated with the page that has been
 309                  * consumed.  See the routine restore_reserve_on_error
 310                  * for details.  Unfortunately, we can not call
 311                  * restore_reserve_on_error now as it would require holding
 312                  * mmap_sem.  Clear the PagePrivate flag so that the global
 313                  * reserve count will not be incremented in free_huge_page.
 314                  * The reservation map will still indicate the reservation
 315                  * was consumed and possibly prevent later page allocation.
 316                  * This is better than leaking a global reservation.
 317                  */
 318                 ClearPagePrivate(page);
 319                 put_page(page);
 320         }
 321         BUG_ON(copied < 0);
 322         BUG_ON(err > 0);
 323         BUG_ON(!copied && !err);
 324         return copied ? copied : err;
 325 }
 326 #else /* !CONFIG_HUGETLB_PAGE */
 327 /* fail at build time if gcc attempts to use this */
 328 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 329                                       struct vm_area_struct *dst_vma,
 330                                       unsigned long dst_start,
 331                                       unsigned long src_start,
 332                                       unsigned long len,
 333                                       bool zeropage);
 334 #endif /* CONFIG_HUGETLB_PAGE */
 335
 336 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 337                                               unsigned long dst_start,
 338                                               unsigned long src_start,
 339                                               unsigned long len,
 340                                               bool zeropage)
 341 {
 342         struct vm_area_struct *dst_vma;
 343         ssize_t err;
 344         pmd_t *dst_pmd;
 345         unsigned long src_addr, dst_addr;
 346         long copied;
 347         struct page *page;
 348
 349         /*
 350          * Sanitize the command parameters:
 351          */
 352         BUG_ON(dst_start & ~PAGE_MASK);
 353         BUG_ON(len & ~PAGE_MASK);
 354
 355         /* Does the address range wrap, or is the span zero-sized? */
 356         BUG_ON(src_start + len <= src_start);
 357         BUG_ON(dst_start + len <= dst_start);
 358
 359         src_addr = src_start;
 360         dst_addr = dst_start;
 361         copied = 0;
 362         page = NULL;
 363 retry:
 364         down_read(&dst_mm->mmap_sem);
 365
 366         /*
 367          * Make sure the vma is not shared, that the dst range is
 368          * both valid and fully within a single existing vma.
 369          */
 370         err = -EINVAL;
 371         dst_vma = find_vma(dst_mm, dst_start);
 372         if (!dst_vma || (dst_vma->vm_flags & VM_SHARED))
 373                 goto out_unlock;
 374         if (dst_start < dst_vma->vm_start ||
 375             dst_start + len > dst_vma->vm_end)
 376                 goto out_unlock;
 377
 378         /*
 379          * If this is a HUGETLB vma, pass off to appropriate routine
 380          */
 381         if (is_vm_hugetlb_page(dst_vma))
 382                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 383                                                 src_start, len, zeropage);
 384
 385         /*
 386          * Be strict and only allow __mcopy_atomic on userfaultfd
 387          * registered ranges to prevent userland errors going
 388          * unnoticed. As far as the VM consistency is concerned, it
 389          * would be perfectly safe to remove this check, but there's
 390          * no useful usage for __mcopy_atomic ouside of userfaultfd
 391          * registered ranges. This is after all why these are ioctls
 392          * belonging to the userfaultfd and not syscalls.
 393          */
 394         if (!dst_vma->vm_userfaultfd_ctx.ctx)
 395                 goto out_unlock;
 396
 397         /*
 398          * FIXME: only allow copying on anonymous vmas, tmpfs should
 399          * be added.
 400          */
 401         if (!vma_is_anonymous(dst_vma))
 402                 goto out_unlock;
 403
 404         /*
 405          * Ensure the dst_vma has a anon_vma or this page
 406          * would get a NULL anon_vma when moved in the
 407          * dst_vma.
 408          */
 409         err = -ENOMEM;
 410         if (unlikely(anon_vma_prepare(dst_vma)))
 411                 goto out_unlock;
 412
 413         while (src_addr < src_start + len) {
 414                 pmd_t dst_pmdval;
 415
 416                 BUG_ON(dst_addr >= dst_start + len);
 417
 418                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 419                 if (unlikely(!dst_pmd)) {
 420                         err = -ENOMEM;
 421                         break;
 422                 }
 423
 424                 dst_pmdval = pmd_read_atomic(dst_pmd);
 425                 /*
 426                  * If the dst_pmd is mapped as THP don't
 427                  * override it and just be strict.
 428                  */
 429                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 430                         err = -EEXIST;
 431                         break;
 432                 }
 433                 if (unlikely(pmd_none(dst_pmdval)) &&
 434                     unlikely(__pte_alloc(dst_mm, dst_pmd, dst_addr))) {
 435                         err = -ENOMEM;
 436                         break;
 437                 }
 438                 /* If an huge pmd materialized from under us fail */
 439                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 440                         err = -EFAULT;
 441                         break;
 442                 }
 443
 444                 BUG_ON(pmd_none(*dst_pmd));
 445                 BUG_ON(pmd_trans_huge(*dst_pmd));
 446
 447                 if (!zeropage)
 448                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 449                                                dst_addr, src_addr, &page);
 450                 else
 451                         err = mfill_zeropage_pte(dst_mm, dst_pmd, dst_vma,
 452                                                  dst_addr);
 453
 454                 cond_resched();
 455
 456                 if (unlikely(err == -EFAULT)) {
 457                         void *page_kaddr;
 458
 459                         up_read(&dst_mm->mmap_sem);
 460                         BUG_ON(!page);
 461
 462                         page_kaddr = kmap(page);
 463                         err = copy_from_user(page_kaddr,
 464                                              (const void __user *) src_addr,
 465                                              PAGE_SIZE);
 466                         kunmap(page);
 467                         if (unlikely(err)) {
 468                                 err = -EFAULT;
 469                                 goto out;
 470                         }
 471                         goto retry;
 472                 } else
 473                         BUG_ON(page);
 474
 475                 if (!err) {
 476                         dst_addr += PAGE_SIZE;
 477                         src_addr += PAGE_SIZE;
 478                         copied += PAGE_SIZE;
 479
 480                         if (fatal_signal_pending(current))
 481                                 err = -EINTR;
 482                 }
 483                 if (err)
 484                         break;
 485         }
 486
 487 out_unlock:
 488         up_read(&dst_mm->mmap_sem);
 489 out:
 490         if (page)
 491                 put_page(page);
 492         BUG_ON(copied < 0);
 493         BUG_ON(err > 0);
 494         BUG_ON(!copied && !err);
 495         return copied ? copied : err;
 496 }
 497
 498 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 499                      unsigned long src_start, unsigned long len)
 500 {
 501         return __mcopy_atomic(dst_mm, dst_start, src_start, len, false);
 502 }
 503
 504 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 505                        unsigned long len)
 506 {
 507         return __mcopy_atomic(dst_mm, start, 0, len, true);
 508 }