linux/mm/gup.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-only
   2#include <linux/kernel.h>
   3#include <linux/errno.h>
   4#include <linux/err.h>
   5#include <linux/spinlock.h>
   6
   7#include <linux/mm.h>
   8#include <linux/memremap.h>
   9#include <linux/pagemap.h>
  10#include <linux/rmap.h>
  11#include <linux/swap.h>
  12#include <linux/swapops.h>
  13
  14#include <linux/sched/signal.h>
  15#include <linux/rwsem.h>
  16#include <linux/hugetlb.h>
  17#include <linux/migrate.h>
  18#include <linux/mm_inline.h>
  19#include <linux/sched/mm.h>
  20
  21#include <asm/mmu_context.h>
  22#include <asm/pgtable.h>
  23#include <asm/tlbflush.h>
  24
  25#include "internal.h"
  26
  27struct follow_page_context {
  28        struct dev_pagemap *pgmap;
  29        unsigned int page_mask;
  30};
  31
  32/**
  33 * put_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
  34 * @pages:  array of pages to be maybe marked dirty, and definitely released.
  35 * @npages: number of pages in the @pages array.
  36 * @make_dirty: whether to mark the pages dirty
  37 *
  38 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
  39 * variants called on that page.
  40 *
  41 * For each page in the @pages array, make that page (or its head page, if a
  42 * compound page) dirty, if @make_dirty is true, and if the page was previously
  43 * listed as clean. In any case, releases all pages using put_user_page(),
  44 * possibly via put_user_pages(), for the non-dirty case.
  45 *
  46 * Please see the put_user_page() documentation for details.
  47 *
  48 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
  49 * required, then the caller should a) verify that this is really correct,
  50 * because _lock() is usually required, and b) hand code it:
  51 * set_page_dirty_lock(), put_user_page().
  52 *
  53 */
  54void put_user_pages_dirty_lock(struct page **pages, unsigned long npages,
  55                               bool make_dirty)
  56{
  57        unsigned long index;
  58
  59        /*
  60         * TODO: this can be optimized for huge pages: if a series of pages is
  61         * physically contiguous and part of the same compound page, then a
  62         * single operation to the head page should suffice.
  63         */
  64
  65        if (!make_dirty) {
  66                put_user_pages(pages, npages);
  67                return;
  68        }
  69
  70        for (index = 0; index < npages; index++) {
  71                struct page *page = compound_head(pages[index]);
  72                /*
  73                 * Checking PageDirty at this point may race with
  74                 * clear_page_dirty_for_io(), but that's OK. Two key
  75                 * cases:
  76                 *
  77                 * 1) This code sees the page as already dirty, so it
  78                 * skips the call to set_page_dirty(). That could happen
  79                 * because clear_page_dirty_for_io() called
  80                 * page_mkclean(), followed by set_page_dirty().
  81                 * However, now the page is going to get written back,
  82                 * which meets the original intention of setting it
  83                 * dirty, so all is well: clear_page_dirty_for_io() goes
  84                 * on to call TestClearPageDirty(), and write the page
  85                 * back.
  86                 *
  87                 * 2) This code sees the page as clean, so it calls
  88                 * set_page_dirty(). The page stays dirty, despite being
  89                 * written back, so it gets written back again in the
  90                 * next writeback cycle. This is harmless.
  91                 */
  92                if (!PageDirty(page))
  93                        set_page_dirty_lock(page);
  94                put_user_page(page);
  95        }
  96}
  97EXPORT_SYMBOL(put_user_pages_dirty_lock);
  98
  99/**
 100 * put_user_pages() - release an array of gup-pinned pages.
 101 * @pages:  array of pages to be marked dirty and released.
 102 * @npages: number of pages in the @pages array.
 103 *
 104 * For each page in the @pages array, release the page using put_user_page().
 105 *
 106 * Please see the put_user_page() documentation for details.
 107 */
 108void put_user_pages(struct page **pages, unsigned long npages)
 109{
 110        unsigned long index;
 111
 112        /*
 113         * TODO: this can be optimized for huge pages: if a series of pages is
 114         * physically contiguous and part of the same compound page, then a
 115         * single operation to the head page should suffice.
 116         */
 117        for (index = 0; index < npages; index++)
 118                put_user_page(pages[index]);
 119}
 120EXPORT_SYMBOL(put_user_pages);
 121
 122#ifdef CONFIG_MMU
 123static struct page *no_page_table(struct vm_area_struct *vma,
 124                unsigned int flags)
 125{
 126        /*
 127         * When core dumping an enormous anonymous area that nobody
 128         * has touched so far, we don't want to allocate unnecessary pages or
 129         * page tables.  Return error instead of NULL to skip handle_mm_fault,
 130         * then get_dump_page() will return NULL to leave a hole in the dump.
 131         * But we can only make this optimization where a hole would surely
 132         * be zero-filled if handle_mm_fault() actually did handle it.
 133         */
 134        if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
 135                return ERR_PTR(-EFAULT);
 136        return NULL;
 137}
 138
 139static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
 140                pte_t *pte, unsigned int flags)
 141{
 142        /* No page to get reference */
 143        if (flags & FOLL_GET)
 144                return -EFAULT;
 145
 146        if (flags & FOLL_TOUCH) {
 147                pte_t entry = *pte;
 148
 149                if (flags & FOLL_WRITE)
 150                        entry = pte_mkdirty(entry);
 151                entry = pte_mkyoung(entry);
 152
 153                if (!pte_same(*pte, entry)) {
 154                        set_pte_at(vma->vm_mm, address, pte, entry);
 155                        update_mmu_cache(vma, address, pte);
 156                }
 157        }
 158
 159        /* Proper page table entry exists, but no corresponding struct page */
 160        return -EEXIST;
 161}
 162
 163/*
 164 * FOLL_FORCE can write to even unwritable pte's, but only
 165 * after we've gone through a COW cycle and they are dirty.
 166 */
 167static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
 168{
 169        return pte_write(pte) ||
 170                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
 171}
 172
 173static struct page *follow_page_pte(struct vm_area_struct *vma,
 174                unsigned long address, pmd_t *pmd, unsigned int flags,
 175                struct dev_pagemap **pgmap)
 176{
 177        struct mm_struct *mm = vma->vm_mm;
 178        struct page *page;
 179        spinlock_t *ptl;
 180        pte_t *ptep, pte;
 181
 182retry:
 183        if (unlikely(pmd_bad(*pmd)))
 184                return no_page_table(vma, flags);
 185
 186        ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 187        pte = *ptep;
 188        if (!pte_present(pte)) {
 189                swp_entry_t entry;
 190                /*
 191                 * KSM's break_ksm() relies upon recognizing a ksm page
 192                 * even while it is being migrated, so for that case we
 193                 * need migration_entry_wait().
 194                 */
 195                if (likely(!(flags & FOLL_MIGRATION)))
 196                        goto no_page;
 197                if (pte_none(pte))
 198                        goto no_page;
 199                entry = pte_to_swp_entry(pte);
 200                if (!is_migration_entry(entry))
 201                        goto no_page;
 202                pte_unmap_unlock(ptep, ptl);
 203                migration_entry_wait(mm, pmd, address);
 204                goto retry;
 205        }
 206        if ((flags & FOLL_NUMA) && pte_protnone(pte))
 207                goto no_page;
 208        if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
 209                pte_unmap_unlock(ptep, ptl);
 210                return NULL;
 211        }
 212
 213        page = vm_normal_page(vma, address, pte);
 214        if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
 215                /*
 216                 * Only return device mapping pages in the FOLL_GET case since
 217                 * they are only valid while holding the pgmap reference.
 218                 */
 219                *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
 220                if (*pgmap)
 221                        page = pte_page(pte);
 222                else
 223                        goto no_page;
 224        } else if (unlikely(!page)) {
 225                if (flags & FOLL_DUMP) {
 226                        /* Avoid special (like zero) pages in core dumps */
 227                        page = ERR_PTR(-EFAULT);
 228                        goto out;
 229                }
 230
 231                if (is_zero_pfn(pte_pfn(pte))) {
 232                        page = pte_page(pte);
 233                } else {
 234                        int ret;
 235
 236                        ret = follow_pfn_pte(vma, address, ptep, flags);
 237                        page = ERR_PTR(ret);
 238                        goto out;
 239                }
 240        }
 241
 242        if (flags & FOLL_SPLIT && PageTransCompound(page)) {
 243                int ret;
 244                get_page(page);
 245                pte_unmap_unlock(ptep, ptl);
 246                lock_page(page);
 247                ret = split_huge_page(page);
 248                unlock_page(page);
 249                put_page(page);
 250                if (ret)
 251                        return ERR_PTR(ret);
 252                goto retry;
 253        }
 254
 255        if (flags & FOLL_GET) {
 256                if (unlikely(!try_get_page(page))) {
 257                        page = ERR_PTR(-ENOMEM);
 258                        goto out;
 259                }
 260        }
 261        if (flags & FOLL_TOUCH) {
 262                if ((flags & FOLL_WRITE) &&
 263                    !pte_dirty(pte) && !PageDirty(page))
 264                        set_page_dirty(page);
 265                /*
 266                 * pte_mkyoung() would be more correct here, but atomic care
 267                 * is needed to avoid losing the dirty bit: it is easier to use
 268                 * mark_page_accessed().
 269                 */
 270                mark_page_accessed(page);
 271        }
 272        if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
 273                /* Do not mlock pte-mapped THP */
 274                if (PageTransCompound(page))
 275                        goto out;
 276
 277                /*
 278                 * The preliminary mapping check is mainly to avoid the
 279                 * pointless overhead of lock_page on the ZERO_PAGE
 280                 * which might bounce very badly if there is contention.
 281                 *
 282                 * If the page is already locked, we don't need to
 283                 * handle it now - vmscan will handle it later if and
 284                 * when it attempts to reclaim the page.
 285                 */
 286                if (page->mapping && trylock_page(page)) {
 287                        lru_add_drain();  /* push cached pages to LRU */
 288                        /*
 289                         * Because we lock page here, and migration is
 290                         * blocked by the pte's page reference, and we
 291                         * know the page is still mapped, we don't even
 292                         * need to check for file-cache page truncation.
 293                         */
 294                        mlock_vma_page(page);
 295                        unlock_page(page);
 296                }
 297        }
 298out:
 299        pte_unmap_unlock(ptep, ptl);
 300        return page;
 301no_page:
 302        pte_unmap_unlock(ptep, ptl);
 303        if (!pte_none(pte))
 304                return NULL;
 305        return no_page_table(vma, flags);
 306}
 307
 308static struct page *follow_pmd_mask(struct vm_area_struct *vma,
 309                                    unsigned long address, pud_t *pudp,
 310                                    unsigned int flags,
 311                                    struct follow_page_context *ctx)
 312{
 313        pmd_t *pmd, pmdval;
 314        spinlock_t *ptl;
 315        struct page *page;
 316        struct mm_struct *mm = vma->vm_mm;
 317
 318        pmd = pmd_offset(pudp, address);
 319        /*
 320         * The READ_ONCE() will stabilize the pmdval in a register or
 321         * on the stack so that it will stop changing under the code.
 322         */
 323        pmdval = READ_ONCE(*pmd);
 324        if (pmd_none(pmdval))
 325                return no_page_table(vma, flags);
 326        if (pmd_huge(pmdval) && vma->vm_flags & VM_HUGETLB) {
 327                page = follow_huge_pmd(mm, address, pmd, flags);
 328                if (page)
 329                        return page;
 330                return no_page_table(vma, flags);
 331        }
 332        if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
 333                page = follow_huge_pd(vma, address,
 334                                      __hugepd(pmd_val(pmdval)), flags,
 335                                      PMD_SHIFT);
 336                if (page)
 337                        return page;
 338                return no_page_table(vma, flags);
 339        }
 340retry:
 341        if (!pmd_present(pmdval)) {
 342                if (likely(!(flags & FOLL_MIGRATION)))
 343                        return no_page_table(vma, flags);
 344                VM_BUG_ON(thp_migration_supported() &&
 345                                  !is_pmd_migration_entry(pmdval));
 346                if (is_pmd_migration_entry(pmdval))
 347                        pmd_migration_entry_wait(mm, pmd);
 348                pmdval = READ_ONCE(*pmd);
 349                /*
 350                 * MADV_DONTNEED may convert the pmd to null because
 351                 * mmap_sem is held in read mode
 352                 */
 353                if (pmd_none(pmdval))
 354                        return no_page_table(vma, flags);
 355                goto retry;
 356        }
 357        if (pmd_devmap(pmdval)) {
 358                ptl = pmd_lock(mm, pmd);
 359                page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
 360                spin_unlock(ptl);
 361                if (page)
 362                        return page;
 363        }
 364        if (likely(!pmd_trans_huge(pmdval)))
 365                return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
 366
 367        if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
 368                return no_page_table(vma, flags);
 369
 370retry_locked:
 371        ptl = pmd_lock(mm, pmd);
 372        if (unlikely(pmd_none(*pmd))) {
 373                spin_unlock(ptl);
 374                return no_page_table(vma, flags);
 375        }
 376        if (unlikely(!pmd_present(*pmd))) {
 377                spin_unlock(ptl);
 378                if (likely(!(flags & FOLL_MIGRATION)))
 379                        return no_page_table(vma, flags);
 380                pmd_migration_entry_wait(mm, pmd);
 381                goto retry_locked;
 382        }
 383        if (unlikely(!pmd_trans_huge(*pmd))) {
 384                spin_unlock(ptl);
 385                return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
 386        }
 387        if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
 388                int ret;
 389                page = pmd_page(*pmd);
 390                if (is_huge_zero_page(page)) {
 391                        spin_unlock(ptl);
 392                        ret = 0;
 393                        split_huge_pmd(vma, pmd, address);
 394                        if (pmd_trans_unstable(pmd))
 395                                ret = -EBUSY;
 396                } else if (flags & FOLL_SPLIT) {
 397                        if (unlikely(!try_get_page(page))) {
 398                                spin_unlock(ptl);
 399                                return ERR_PTR(-ENOMEM);
 400                        }
 401                        spin_unlock(ptl);
 402                        lock_page(page);
 403                        ret = split_huge_page(page);
 404                        unlock_page(page);
 405                        put_page(page);
 406                        if (pmd_none(*pmd))
 407                                return no_page_table(vma, flags);
 408                } else {  /* flags & FOLL_SPLIT_PMD */
 409                        spin_unlock(ptl);
 410                        split_huge_pmd(vma, pmd, address);
 411                        ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
 412                }
 413
 414                return ret ? ERR_PTR(ret) :
 415                        follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
 416        }
 417        page = follow_trans_huge_pmd(vma, address, pmd, flags);
 418        spin_unlock(ptl);
 419        ctx->page_mask = HPAGE_PMD_NR - 1;
 420        return page;
 421}
 422
 423static struct page *follow_pud_mask(struct vm_area_struct *vma,
 424                                    unsigned long address, p4d_t *p4dp,
 425                                    unsigned int flags,
 426                                    struct follow_page_context *ctx)
 427{
 428        pud_t *pud;
 429        spinlock_t *ptl;
 430        struct page *page;
 431        struct mm_struct *mm = vma->vm_mm;
 432
 433        pud = pud_offset(p4dp, address);
 434        if (pud_none(*pud))
 435                return no_page_table(vma, flags);
 436        if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
 437                page = follow_huge_pud(mm, address, pud, flags);
 438                if (page)
 439                        return page;
 440                return no_page_table(vma, flags);
 441        }
 442        if (is_hugepd(__hugepd(pud_val(*pud)))) {
 443                page = follow_huge_pd(vma, address,
 444                                      __hugepd(pud_val(*pud)), flags,
 445                                      PUD_SHIFT);
 446                if (page)
 447                        return page;
 448                return no_page_table(vma, flags);
 449        }
 450        if (pud_devmap(*pud)) {
 451                ptl = pud_lock(mm, pud);
 452                page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
 453                spin_unlock(ptl);
 454                if (page)
 455                        return page;
 456        }
 457        if (unlikely(pud_bad(*pud)))
 458                return no_page_table(vma, flags);
 459
 460        return follow_pmd_mask(vma, address, pud, flags, ctx);
 461}
 462
 463static struct page *follow_p4d_mask(struct vm_area_struct *vma,
 464                                    unsigned long address, pgd_t *pgdp,
 465                                    unsigned int flags,
 466                                    struct follow_page_context *ctx)
 467{
 468        p4d_t *p4d;
 469        struct page *page;
 470
 471        p4d = p4d_offset(pgdp, address);
 472        if (p4d_none(*p4d))
 473                return no_page_table(vma, flags);
 474        BUILD_BUG_ON(p4d_huge(*p4d));
 475        if (unlikely(p4d_bad(*p4d)))
 476                return no_page_table(vma, flags);
 477
 478        if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
 479                page = follow_huge_pd(vma, address,
 480                                      __hugepd(p4d_val(*p4d)), flags,
 481                                      P4D_SHIFT);
 482                if (page)
 483                        return page;
 484                return no_page_table(vma, flags);
 485        }
 486        return follow_pud_mask(vma, address, p4d, flags, ctx);
 487}
 488
 489/**
 490 * follow_page_mask - look up a page descriptor from a user-virtual address
 491 * @vma: vm_area_struct mapping @address
 492 * @address: virtual address to look up
 493 * @flags: flags modifying lookup behaviour
 494 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
 495 *       pointer to output page_mask
 496 *
 497 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
 498 *
 499 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
 500 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
 501 *
 502 * On output, the @ctx->page_mask is set according to the size of the page.
 503 *
 504 * Return: the mapped (struct page *), %NULL if no mapping exists, or
 505 * an error pointer if there is a mapping to something not represented
 506 * by a page descriptor (see also vm_normal_page()).
 507 */
 508static struct page *follow_page_mask(struct vm_area_struct *vma,
 509                              unsigned long address, unsigned int flags,
 510                              struct follow_page_context *ctx)
 511{
 512        pgd_t *pgd;
 513        struct page *page;
 514        struct mm_struct *mm = vma->vm_mm;
 515
 516        ctx->page_mask = 0;
 517
 518        /* make this handle hugepd */
 519        page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
 520        if (!IS_ERR(page)) {
 521                BUG_ON(flags & FOLL_GET);
 522                return page;
 523        }
 524
 525        pgd = pgd_offset(mm, address);
 526
 527        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 528                return no_page_table(vma, flags);
 529
 530        if (pgd_huge(*pgd)) {
 531                page = follow_huge_pgd(mm, address, pgd, flags);
 532                if (page)
 533                        return page;
 534                return no_page_table(vma, flags);
 535        }
 536        if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
 537                page = follow_huge_pd(vma, address,
 538                                      __hugepd(pgd_val(*pgd)), flags,
 539                                      PGDIR_SHIFT);
 540                if (page)
 541                        return page;
 542                return no_page_table(vma, flags);
 543        }
 544
 545        return follow_p4d_mask(vma, address, pgd, flags, ctx);
 546}
 547
 548struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
 549                         unsigned int foll_flags)
 550{
 551        struct follow_page_context ctx = { NULL };
 552        struct page *page;
 553
 554        page = follow_page_mask(vma, address, foll_flags, &ctx);
 555        if (ctx.pgmap)
 556                put_dev_pagemap(ctx.pgmap);
 557        return page;
 558}
 559
 560static int get_gate_page(struct mm_struct *mm, unsigned long address,
 561                unsigned int gup_flags, struct vm_area_struct **vma,
 562                struct page **page)
 563{
 564        pgd_t *pgd;
 565        p4d_t *p4d;
 566        pud_t *pud;
 567        pmd_t *pmd;
 568        pte_t *pte;
 569        int ret = -EFAULT;
 570
 571        /* user gate pages are read-only */
 572        if (gup_flags & FOLL_WRITE)
 573                return -EFAULT;
 574        if (address > TASK_SIZE)
 575                pgd = pgd_offset_k(address);
 576        else
 577                pgd = pgd_offset_gate(mm, address);
 578        if (pgd_none(*pgd))
 579                return -EFAULT;
 580        p4d = p4d_offset(pgd, address);
 581        if (p4d_none(*p4d))
 582                return -EFAULT;
 583        pud = pud_offset(p4d, address);
 584        if (pud_none(*pud))
 585                return -EFAULT;
 586        pmd = pmd_offset(pud, address);
 587        if (!pmd_present(*pmd))
 588                return -EFAULT;
 589        VM_BUG_ON(pmd_trans_huge(*pmd));
 590        pte = pte_offset_map(pmd, address);
 591        if (pte_none(*pte))
 592                goto unmap;
 593        *vma = get_gate_vma(mm);
 594        if (!page)
 595                goto out;
 596        *page = vm_normal_page(*vma, address, *pte);
 597        if (!*page) {
 598                if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
 599                        goto unmap;
 600                *page = pte_page(*pte);
 601        }
 602        if (unlikely(!try_get_page(*page))) {
 603                ret = -ENOMEM;
 604                goto unmap;
 605        }
 606out:
 607        ret = 0;
 608unmap:
 609        pte_unmap(pte);
 610        return ret;
 611}
 612
 613/*
 614 * mmap_sem must be held on entry.  If @nonblocking != NULL and
 615 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
 616 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
 617 */
 618static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
 619                unsigned long address, unsigned int *flags, int *nonblocking)
 620{
 621        unsigned int fault_flags = 0;
 622        vm_fault_t ret;
 623
 624        /* mlock all present pages, but do not fault in new pages */
 625        if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
 626                return -ENOENT;
 627        if (*flags & FOLL_WRITE)
 628                fault_flags |= FAULT_FLAG_WRITE;
 629        if (*flags & FOLL_REMOTE)
 630                fault_flags |= FAULT_FLAG_REMOTE;
 631        if (nonblocking)
 632                fault_flags |= FAULT_FLAG_ALLOW_RETRY;
 633        if (*flags & FOLL_NOWAIT)
 634                fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
 635        if (*flags & FOLL_TRIED) {
 636                VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
 637                fault_flags |= FAULT_FLAG_TRIED;
 638        }
 639
 640        ret = handle_mm_fault(vma, address, fault_flags);
 641        if (ret & VM_FAULT_ERROR) {
 642                int err = vm_fault_to_errno(ret, *flags);
 643
 644                if (err)
 645                        return err;
 646                BUG();
 647        }
 648
 649        if (tsk) {
 650                if (ret & VM_FAULT_MAJOR)
 651                        tsk->maj_flt++;
 652                else
 653                        tsk->min_flt++;
 654        }
 655
 656        if (ret & VM_FAULT_RETRY) {
 657                if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
 658                        *nonblocking = 0;
 659                return -EBUSY;
 660        }
 661
 662        /*
 663         * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
 664         * necessary, even if maybe_mkwrite decided not to set pte_write. We
 665         * can thus safely do subsequent page lookups as if they were reads.
 666         * But only do so when looping for pte_write is futile: in some cases
 667         * userspace may also be wanting to write to the gotten user page,
 668         * which a read fault here might prevent (a readonly page might get
 669         * reCOWed by userspace write).
 670         */
 671        if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
 672                *flags |= FOLL_COW;
 673        return 0;
 674}
 675
 676static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
 677{
 678        vm_flags_t vm_flags = vma->vm_flags;
 679        int write = (gup_flags & FOLL_WRITE);
 680        int foreign = (gup_flags & FOLL_REMOTE);
 681
 682        if (vm_flags & (VM_IO | VM_PFNMAP))
 683                return -EFAULT;
 684
 685        if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
 686                return -EFAULT;
 687
 688        if (write) {
 689                if (!(vm_flags & VM_WRITE)) {
 690                        if (!(gup_flags & FOLL_FORCE))
 691                                return -EFAULT;
 692                        /*
 693                         * We used to let the write,force case do COW in a
 694                         * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
 695                         * set a breakpoint in a read-only mapping of an
 696                         * executable, without corrupting the file (yet only
 697                         * when that file had been opened for writing!).
 698                         * Anon pages in shared mappings are surprising: now
 699                         * just reject it.
 700                         */
 701                        if (!is_cow_mapping(vm_flags))
 702                                return -EFAULT;
 703                }
 704        } else if (!(vm_flags & VM_READ)) {
 705                if (!(gup_flags & FOLL_FORCE))
 706                        return -EFAULT;
 707                /*
 708                 * Is there actually any vma we can reach here which does not
 709                 * have VM_MAYREAD set?
 710                 */
 711                if (!(vm_flags & VM_MAYREAD))
 712                        return -EFAULT;
 713        }
 714        /*
 715         * gups are always data accesses, not instruction
 716         * fetches, so execute=false here
 717         */
 718        if (!arch_vma_access_permitted(vma, write, false, foreign))
 719                return -EFAULT;
 720        return 0;
 721}
 722
 723/**
 724 * __get_user_pages() - pin user pages in memory
 725 * @tsk:        task_struct of target task
 726 * @mm:         mm_struct of target mm
 727 * @start:      starting user address
 728 * @nr_pages:   number of pages from start to pin
 729 * @gup_flags:  flags modifying pin behaviour
 730 * @pages:      array that receives pointers to the pages pinned.
 731 *              Should be at least nr_pages long. Or NULL, if caller
 732 *              only intends to ensure the pages are faulted in.
 733 * @vmas:       array of pointers to vmas corresponding to each page.
 734 *              Or NULL if the caller does not require them.
 735 * @nonblocking: whether waiting for disk IO or mmap_sem contention
 736 *
 737 * Returns number of pages pinned. This may be fewer than the number
 738 * requested. If nr_pages is 0 or negative, returns 0. If no pages
 739 * were pinned, returns -errno. Each page returned must be released
 740 * with a put_page() call when it is finished with. vmas will only
 741 * remain valid while mmap_sem is held.
 742 *
 743 * Must be called with mmap_sem held.  It may be released.  See below.
 744 *
 745 * __get_user_pages walks a process's page tables and takes a reference to
 746 * each struct page that each user address corresponds to at a given
 747 * instant. That is, it takes the page that would be accessed if a user
 748 * thread accesses the given user virtual address at that instant.
 749 *
 750 * This does not guarantee that the page exists in the user mappings when
 751 * __get_user_pages returns, and there may even be a completely different
 752 * page there in some cases (eg. if mmapped pagecache has been invalidated
 753 * and subsequently re faulted). However it does guarantee that the page
 754 * won't be freed completely. And mostly callers simply care that the page
 755 * contains data that was valid *at some point in time*. Typically, an IO
 756 * or similar operation cannot guarantee anything stronger anyway because
 757 * locks can't be held over the syscall boundary.
 758 *
 759 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
 760 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
 761 * appropriate) must be called after the page is finished with, and
 762 * before put_page is called.
 763 *
 764 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
 765 * or mmap_sem contention, and if waiting is needed to pin all pages,
 766 * *@nonblocking will be set to 0.  Further, if @gup_flags does not
 767 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
 768 * this case.
 769 *
 770 * A caller using such a combination of @nonblocking and @gup_flags
 771 * must therefore hold the mmap_sem for reading only, and recognize
 772 * when it's been released.  Otherwise, it must be held for either
 773 * reading or writing and will not be released.
 774 *
 775 * In most cases, get_user_pages or get_user_pages_fast should be used
 776 * instead of __get_user_pages. __get_user_pages should be used only if
 777 * you need some special @gup_flags.
 778 */
 779static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 780                unsigned long start, unsigned long nr_pages,
 781                unsigned int gup_flags, struct page **pages,
 782                struct vm_area_struct **vmas, int *nonblocking)
 783{
 784        long ret = 0, i = 0;
 785        struct vm_area_struct *vma = NULL;
 786        struct follow_page_context ctx = { NULL };
 787
 788        if (!nr_pages)
 789                return 0;
 790
 791        start = untagged_addr(start);
 792
 793        VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
 794
 795        /*
 796         * If FOLL_FORCE is set then do not force a full fault as the hinting
 797         * fault information is unrelated to the reference behaviour of a task
 798         * using the address space
 799         */
 800        if (!(gup_flags & FOLL_FORCE))
 801                gup_flags |= FOLL_NUMA;
 802
 803        do {
 804                struct page *page;
 805                unsigned int foll_flags = gup_flags;
 806                unsigned int page_increm;
 807
 808                /* first iteration or cross vma bound */
 809                if (!vma || start >= vma->vm_end) {
 810                        vma = find_extend_vma(mm, start);
 811                        if (!vma && in_gate_area(mm, start)) {
 812                                ret = get_gate_page(mm, start & PAGE_MASK,
 813                                                gup_flags, &vma,
 814                                                pages ? &pages[i] : NULL);
 815                                if (ret)
 816                                        goto out;
 817                                ctx.page_mask = 0;
 818                                goto next_page;
 819                        }
 820
 821                        if (!vma || check_vma_flags(vma, gup_flags)) {
 822                                ret = -EFAULT;
 823                                goto out;
 824                        }
 825                        if (is_vm_hugetlb_page(vma)) {
 826                                i = follow_hugetlb_page(mm, vma, pages, vmas,
 827                                                &start, &nr_pages, i,
 828                                                gup_flags, nonblocking);
 829                                continue;
 830                        }
 831                }
 832retry:
 833                /*
 834                 * If we have a pending SIGKILL, don't keep faulting pages and
 835                 * potentially allocating memory.
 836                 */
 837                if (fatal_signal_pending(current)) {
 838                        ret = -ERESTARTSYS;
 839                        goto out;
 840                }
 841                cond_resched();
 842
 843                page = follow_page_mask(vma, start, foll_flags, &ctx);
 844                if (!page) {
 845                        ret = faultin_page(tsk, vma, start, &foll_flags,
 846                                        nonblocking);
 847                        switch (ret) {
 848                        case 0:
 849                                goto retry;
 850                        case -EBUSY:
 851                                ret = 0;
 852                                /* FALLTHRU */
 853                        case -EFAULT:
 854                        case -ENOMEM:
 855                        case -EHWPOISON:
 856                                goto out;
 857                        case -ENOENT:
 858                                goto next_page;
 859                        }
 860                        BUG();
 861                } else if (PTR_ERR(page) == -EEXIST) {
 862                        /*
 863                         * Proper page table entry exists, but no corresponding
 864                         * struct page.
 865                         */
 866                        goto next_page;
 867                } else if (IS_ERR(page)) {
 868                        ret = PTR_ERR(page);
 869                        goto out;
 870                }
 871                if (pages) {
 872                        pages[i] = page;
 873                        flush_anon_page(vma, page, start);
 874                        flush_dcache_page(page);
 875                        ctx.page_mask = 0;
 876                }
 877next_page:
 878                if (vmas) {
 879                        vmas[i] = vma;
 880                        ctx.page_mask = 0;
 881                }
 882                page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
 883                if (page_increm > nr_pages)
 884                        page_increm = nr_pages;
 885                i += page_increm;
 886                start += page_increm * PAGE_SIZE;
 887                nr_pages -= page_increm;
 888        } while (nr_pages);
 889out:
 890        if (ctx.pgmap)
 891                put_dev_pagemap(ctx.pgmap);
 892        return i ? i : ret;
 893}
 894
 895static bool vma_permits_fault(struct vm_area_struct *vma,
 896                              unsigned int fault_flags)
 897{
 898        bool write   = !!(fault_flags & FAULT_FLAG_WRITE);
 899        bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
 900        vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
 901
 902        if (!(vm_flags & vma->vm_flags))
 903                return false;
 904
 905        /*
 906         * The architecture might have a hardware protection
 907         * mechanism other than read/write that can deny access.
 908         *
 909         * gup always represents data access, not instruction
 910         * fetches, so execute=false here:
 911         */
 912        if (!arch_vma_access_permitted(vma, write, false, foreign))
 913                return false;
 914
 915        return true;
 916}
 917
 918/*
 919 * fixup_user_fault() - manually resolve a user page fault
 920 * @tsk:        the task_struct to use for page fault accounting, or
 921 *              NULL if faults are not to be recorded.
 922 * @mm:         mm_struct of target mm
 923 * @address:    user address
 924 * @fault_flags:flags to pass down to handle_mm_fault()
 925 * @unlocked:   did we unlock the mmap_sem while retrying, maybe NULL if caller
 926 *              does not allow retry
 927 *
 928 * This is meant to be called in the specific scenario where for locking reasons
 929 * we try to access user memory in atomic context (within a pagefault_disable()
 930 * section), this returns -EFAULT, and we want to resolve the user fault before
 931 * trying again.
 932 *
 933 * Typically this is meant to be used by the futex code.
 934 *
 935 * The main difference with get_user_pages() is that this function will
 936 * unconditionally call handle_mm_fault() which will in turn perform all the
 937 * necessary SW fixup of the dirty and young bits in the PTE, while
 938 * get_user_pages() only guarantees to update these in the struct page.
 939 *
 940 * This is important for some architectures where those bits also gate the
 941 * access permission to the page because they are maintained in software.  On
 942 * such architectures, gup() will not be enough to make a subsequent access
 943 * succeed.
 944 *
 945 * This function will not return with an unlocked mmap_sem. So it has not the
 946 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
 947 */
 948int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
 949                     unsigned long address, unsigned int fault_flags,
 950                     bool *unlocked)
 951{
 952        struct vm_area_struct *vma;
 953        vm_fault_t ret, major = 0;
 954
 955        address = untagged_addr(address);
 956
 957        if (unlocked)
 958                fault_flags |= FAULT_FLAG_ALLOW_RETRY;
 959
 960retry:
 961        vma = find_extend_vma(mm, address);
 962        if (!vma || address < vma->vm_start)
 963                return -EFAULT;
 964
 965        if (!vma_permits_fault(vma, fault_flags))
 966                return -EFAULT;
 967
 968        ret = handle_mm_fault(vma, address, fault_flags);
 969        major |= ret & VM_FAULT_MAJOR;
 970        if (ret & VM_FAULT_ERROR) {
 971                int err = vm_fault_to_errno(ret, 0);
 972
 973                if (err)
 974                        return err;
 975                BUG();
 976        }
 977
 978        if (ret & VM_FAULT_RETRY) {
 979                down_read(&mm->mmap_sem);
 980                if (!(fault_flags & FAULT_FLAG_TRIED)) {
 981                        *unlocked = true;
 982                        fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
 983                        fault_flags |= FAULT_FLAG_TRIED;
 984                        goto retry;
 985                }
 986        }
 987
 988        if (tsk) {
 989                if (major)
 990                        tsk->maj_flt++;
 991                else
 992                        tsk->min_flt++;
 993        }
 994        return 0;
 995}
 996EXPORT_SYMBOL_GPL(fixup_user_fault);
 997
 998static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
 999                                                struct mm_struct *mm,
1000                                                unsigned long start,
1001                                                unsigned long nr_pages,
1002                                                struct page **pages,
1003                                                struct vm_area_struct **vmas,
1004                                                int *locked,
1005                                                unsigned int flags)
1006{
1007        long ret, pages_done;
1008        bool lock_dropped;
1009
1010        if (locked) {
1011                /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1012                BUG_ON(vmas);
1013                /* check caller initialized locked */
1014                BUG_ON(*locked != 1);
1015        }
1016
1017        if (pages)
1018                flags |= FOLL_GET;
1019
1020        pages_done = 0;
1021        lock_dropped = false;
1022        for (;;) {
1023                ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
1024                                       vmas, locked);
1025                if (!locked)
1026                        /* VM_FAULT_RETRY couldn't trigger, bypass */
1027                        return ret;
1028
1029                /* VM_FAULT_RETRY cannot return errors */
1030                if (!*locked) {
1031                        BUG_ON(ret < 0);
1032                        BUG_ON(ret >= nr_pages);
1033                }
1034
1035                if (ret > 0) {
1036                        nr_pages -= ret;
1037                        pages_done += ret;
1038                        if (!nr_pages)
1039                                break;
1040                }
1041                if (*locked) {
1042                        /*
1043                         * VM_FAULT_RETRY didn't trigger or it was a
1044                         * FOLL_NOWAIT.
1045                         */
1046                        if (!pages_done)
1047                                pages_done = ret;
1048                        break;
1049                }
1050                /*
1051                 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1052                 * For the prefault case (!pages) we only update counts.
1053                 */
1054                if (likely(pages))
1055                        pages += ret;
1056                start += ret << PAGE_SHIFT;
1057
1058                /*
1059                 * Repeat on the address that fired VM_FAULT_RETRY
1060                 * without FAULT_FLAG_ALLOW_RETRY but with
1061                 * FAULT_FLAG_TRIED.
1062                 */
1063                *locked = 1;
1064                lock_dropped = true;
1065                down_read(&mm->mmap_sem);
1066                ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
1067                                       pages, NULL, NULL);
1068                if (ret != 1) {
1069                        BUG_ON(ret > 1);
1070                        if (!pages_done)
1071                                pages_done = ret;
1072                        break;
1073                }
1074                nr_pages--;
1075                pages_done++;
1076                if (!nr_pages)
1077                        break;
1078                if (likely(pages))
1079                        pages++;
1080                start += PAGE_SIZE;
1081        }
1082        if (lock_dropped && *locked) {
1083                /*
1084                 * We must let the caller know we temporarily dropped the lock
1085                 * and so the critical section protected by it was lost.
1086                 */
1087                up_read(&mm->mmap_sem);
1088                *locked = 0;
1089        }
1090        return pages_done;
1091}
1092
1093/*
1094 * get_user_pages_remote() - pin user pages in memory
1095 * @tsk:        the task_struct to use for page fault accounting, or
1096 *              NULL if faults are not to be recorded.
1097 * @mm:         mm_struct of target mm
1098 * @start:      starting user address
1099 * @nr_pages:   number of pages from start to pin
1100 * @gup_flags:  flags modifying lookup behaviour
1101 * @pages:      array that receives pointers to the pages pinned.
1102 *              Should be at least nr_pages long. Or NULL, if caller
1103 *              only intends to ensure the pages are faulted in.
1104 * @vmas:       array of pointers to vmas corresponding to each page.
1105 *              Or NULL if the caller does not require them.
1106 * @locked:     pointer to lock flag indicating whether lock is held and
1107 *              subsequently whether VM_FAULT_RETRY functionality can be
1108 *              utilised. Lock must initially be held.
1109 *
1110 * Returns number of pages pinned. This may be fewer than the number
1111 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1112 * were pinned, returns -errno. Each page returned must be released
1113 * with a put_page() call when it is finished with. vmas will only
1114 * remain valid while mmap_sem is held.
1115 *
1116 * Must be called with mmap_sem held for read or write.
1117 *
1118 * get_user_pages walks a process's page tables and takes a reference to
1119 * each struct page that each user address corresponds to at a given
1120 * instant. That is, it takes the page that would be accessed if a user
1121 * thread accesses the given user virtual address at that instant.
1122 *
1123 * This does not guarantee that the page exists in the user mappings when
1124 * get_user_pages returns, and there may even be a completely different
1125 * page there in some cases (eg. if mmapped pagecache has been invalidated
1126 * and subsequently re faulted). However it does guarantee that the page
1127 * won't be freed completely. And mostly callers simply care that the page
1128 * contains data that was valid *at some point in time*. Typically, an IO
1129 * or similar operation cannot guarantee anything stronger anyway because
1130 * locks can't be held over the syscall boundary.
1131 *
1132 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1133 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1134 * be called after the page is finished with, and before put_page is called.
1135 *
1136 * get_user_pages is typically used for fewer-copy IO operations, to get a
1137 * handle on the memory by some means other than accesses via the user virtual
1138 * addresses. The pages may be submitted for DMA to devices or accessed via
1139 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1140 * use the correct cache flushing APIs.
1141 *
1142 * See also get_user_pages_fast, for performance critical applications.
1143 *
1144 * get_user_pages should be phased out in favor of
1145 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1146 * should use get_user_pages because it cannot pass
1147 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1148 */
1149long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1150                unsigned long start, unsigned long nr_pages,
1151                unsigned int gup_flags, struct page **pages,
1152                struct vm_area_struct **vmas, int *locked)
1153{
1154        /*
1155         * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1156         * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1157         * vmas.  As there are no users of this flag in this call we simply
1158         * disallow this option for now.
1159         */
1160        if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1161                return -EINVAL;
1162
1163        return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1164                                       locked,
1165                                       gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1166}
1167EXPORT_SYMBOL(get_user_pages_remote);
1168
1169/**
1170 * populate_vma_page_range() -  populate a range of pages in the vma.
1171 * @vma:   target vma
1172 * @start: start address
1173 * @end:   end address
1174 * @nonblocking:
1175 *
1176 * This takes care of mlocking the pages too if VM_LOCKED is set.
1177 *
1178 * return 0 on success, negative error code on error.
1179 *
1180 * vma->vm_mm->mmap_sem must be held.
1181 *
1182 * If @nonblocking is NULL, it may be held for read or write and will
1183 * be unperturbed.
1184 *
1185 * If @nonblocking is non-NULL, it must held for read only and may be
1186 * released.  If it's released, *@nonblocking will be set to 0.
1187 */
1188long populate_vma_page_range(struct vm_area_struct *vma,
1189                unsigned long start, unsigned long end, int *nonblocking)
1190{
1191        struct mm_struct *mm = vma->vm_mm;
1192        unsigned long nr_pages = (end - start) / PAGE_SIZE;
1193        int gup_flags;
1194
1195        VM_BUG_ON(start & ~PAGE_MASK);
1196        VM_BUG_ON(end   & ~PAGE_MASK);
1197        VM_BUG_ON_VMA(start < vma->vm_start, vma);
1198        VM_BUG_ON_VMA(end   > vma->vm_end, vma);
1199        VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1200
1201        gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1202        if (vma->vm_flags & VM_LOCKONFAULT)
1203                gup_flags &= ~FOLL_POPULATE;
1204        /*
1205         * We want to touch writable mappings with a write fault in order
1206         * to break COW, except for shared mappings because these don't COW
1207         * and we would not want to dirty them for nothing.
1208         */
1209        if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1210                gup_flags |= FOLL_WRITE;
1211
1212        /*
1213         * We want mlock to succeed for regions that have any permissions
1214         * other than PROT_NONE.
1215         */
1216        if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1217                gup_flags |= FOLL_FORCE;
1218
1219        /*
1220         * We made sure addr is within a VMA, so the following will
1221         * not result in a stack expansion that recurses back here.
1222         */
1223        return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1224                                NULL, NULL, nonblocking);
1225}
1226
1227/*
1228 * __mm_populate - populate and/or mlock pages within a range of address space.
1229 *
1230 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1231 * flags. VMAs must be already marked with the desired vm_flags, and
1232 * mmap_sem must not be held.
1233 */
1234int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1235{
1236        struct mm_struct *mm = current->mm;
1237        unsigned long end, nstart, nend;
1238        struct vm_area_struct *vma = NULL;
1239        int locked = 0;
1240        long ret = 0;
1241
1242        end = start + len;
1243
1244        for (nstart = start; nstart < end; nstart = nend) {
1245                /*
1246                 * We want to fault in pages for [nstart; end) address range.
1247                 * Find first corresponding VMA.
1248                 */
1249                if (!locked) {
1250                        locked = 1;
1251                        down_read(&mm->mmap_sem);
1252                        vma = find_vma(mm, nstart);
1253                } else if (nstart >= vma->vm_end)
1254                        vma = vma->vm_next;
1255                if (!vma || vma->vm_start >= end)
1256                        break;
1257                /*
1258                 * Set [nstart; nend) to intersection of desired address
1259                 * range with the first VMA. Also, skip undesirable VMA types.
1260                 */
1261                nend = min(end, vma->vm_end);
1262                if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1263                        continue;
1264                if (nstart < vma->vm_start)
1265                        nstart = vma->vm_start;
1266                /*
1267                 * Now fault in a range of pages. populate_vma_page_range()
1268                 * double checks the vma flags, so that it won't mlock pages
1269                 * if the vma was already munlocked.
1270                 */
1271                ret = populate_vma_page_range(vma, nstart, nend, &locked);
1272                if (ret < 0) {
1273                        if (ignore_errors) {
1274                                ret = 0;
1275                                continue;       /* continue at next VMA */
1276                        }
1277                        break;
1278                }
1279                nend = nstart + ret * PAGE_SIZE;
1280                ret = 0;
1281        }
1282        if (locked)
1283                up_read(&mm->mmap_sem);
1284        return ret;     /* 0 or negative error code */
1285}
1286
1287/**
1288 * get_dump_page() - pin user page in memory while writing it to core dump
1289 * @addr: user address
1290 *
1291 * Returns struct page pointer of user page pinned for dump,
1292 * to be freed afterwards by put_page().
1293 *
1294 * Returns NULL on any kind of failure - a hole must then be inserted into
1295 * the corefile, to preserve alignment with its headers; and also returns
1296 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1297 * allowing a hole to be left in the corefile to save diskspace.
1298 *
1299 * Called without mmap_sem, but after all other threads have been killed.
1300 */
1301#ifdef CONFIG_ELF_CORE
1302struct page *get_dump_page(unsigned long addr)
1303{
1304        struct vm_area_struct *vma;
1305        struct page *page;
1306
1307        if (__get_user_pages(current, current->mm, addr, 1,
1308                             FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1309                             NULL) < 1)
1310                return NULL;
1311        flush_cache_page(vma, addr, page_to_pfn(page));
1312        return page;
1313}
1314#endif /* CONFIG_ELF_CORE */
1315#else /* CONFIG_MMU */
1316static long __get_user_pages_locked(struct task_struct *tsk,
1317                struct mm_struct *mm, unsigned long start,
1318                unsigned long nr_pages, struct page **pages,
1319                struct vm_area_struct **vmas, int *locked,
1320                unsigned int foll_flags)
1321{
1322        struct vm_area_struct *vma;
1323        unsigned long vm_flags;
1324        int i;
1325
1326        /* calculate required read or write permissions.
1327         * If FOLL_FORCE is set, we only require the "MAY" flags.
1328         */
1329        vm_flags  = (foll_flags & FOLL_WRITE) ?
1330                        (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1331        vm_flags &= (foll_flags & FOLL_FORCE) ?
1332                        (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1333
1334        for (i = 0; i < nr_pages; i++) {
1335                vma = find_vma(mm, start);
1336                if (!vma)
1337                        goto finish_or_fault;
1338
1339                /* protect what we can, including chardevs */
1340                if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1341                    !(vm_flags & vma->vm_flags))
1342                        goto finish_or_fault;
1343
1344                if (pages) {
1345                        pages[i] = virt_to_page(start);
1346                        if (pages[i])
1347                                get_page(pages[i]);
1348                }
1349                if (vmas)
1350                        vmas[i] = vma;
1351                start = (start + PAGE_SIZE) & PAGE_MASK;
1352        }
1353
1354        return i;
1355
1356finish_or_fault:
1357        return i ? : -EFAULT;
1358}
1359#endif /* !CONFIG_MMU */
1360
1361#if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1362static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1363{
1364        long i;
1365        struct vm_area_struct *vma_prev = NULL;
1366
1367        for (i = 0; i < nr_pages; i++) {
1368                struct vm_area_struct *vma = vmas[i];
1369
1370                if (vma == vma_prev)
1371                        continue;
1372
1373                vma_prev = vma;
1374
1375                if (vma_is_fsdax(vma))
1376                        return true;
1377        }
1378        return false;
1379}
1380
1381#ifdef CONFIG_CMA
1382static struct page *new_non_cma_page(struct page *page, unsigned long private)
1383{
1384        /*
1385         * We want to make sure we allocate the new page from the same node
1386         * as the source page.
1387         */
1388        int nid = page_to_nid(page);
1389        /*
1390         * Trying to allocate a page for migration. Ignore allocation
1391         * failure warnings. We don't force __GFP_THISNODE here because
1392         * this node here is the node where we have CMA reservation and
1393         * in some case these nodes will have really less non movable
1394         * allocation memory.
1395         */
1396        gfp_t gfp_mask = GFP_USER | __GFP_NOWARN;
1397
1398        if (PageHighMem(page))
1399                gfp_mask |= __GFP_HIGHMEM;
1400
1401#ifdef CONFIG_HUGETLB_PAGE
1402        if (PageHuge(page)) {
1403                struct hstate *h = page_hstate(page);
1404                /*
1405                 * We don't want to dequeue from the pool because pool pages will
1406                 * mostly be from the CMA region.
1407                 */
1408                return alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1409        }
1410#endif
1411        if (PageTransHuge(page)) {
1412                struct page *thp;
1413                /*
1414                 * ignore allocation failure warnings
1415                 */
1416                gfp_t thp_gfpmask = GFP_TRANSHUGE | __GFP_NOWARN;
1417
1418                /*
1419                 * Remove the movable mask so that we don't allocate from
1420                 * CMA area again.
1421                 */
1422                thp_gfpmask &= ~__GFP_MOVABLE;
1423                thp = __alloc_pages_node(nid, thp_gfpmask, HPAGE_PMD_ORDER);
1424                if (!thp)
1425                        return NULL;
1426                prep_transhuge_page(thp);
1427                return thp;
1428        }
1429
1430        return __alloc_pages_node(nid, gfp_mask, 0);
1431}
1432
1433static long check_and_migrate_cma_pages(struct task_struct *tsk,
1434                                        struct mm_struct *mm,
1435                                        unsigned long start,
1436                                        unsigned long nr_pages,
1437                                        struct page **pages,
1438                                        struct vm_area_struct **vmas,
1439                                        unsigned int gup_flags)
1440{
1441        unsigned long i;
1442        unsigned long step;
1443        bool drain_allow = true;
1444        bool migrate_allow = true;
1445        LIST_HEAD(cma_page_list);
1446
1447check_again:
1448        for (i = 0; i < nr_pages;) {
1449
1450                struct page *head = compound_head(pages[i]);
1451
1452                /*
1453                 * gup may start from a tail page. Advance step by the left
1454                 * part.
1455                 */
1456                step = compound_nr(head) - (pages[i] - head);
1457                /*
1458                 * If we get a page from the CMA zone, since we are going to
1459                 * be pinning these entries, we might as well move them out
1460                 * of the CMA zone if possible.
1461                 */
1462                if (is_migrate_cma_page(head)) {
1463                        if (PageHuge(head))
1464                                isolate_huge_page(head, &cma_page_list);
1465                        else {
1466                                if (!PageLRU(head) && drain_allow) {
1467                                        lru_add_drain_all();
1468                                        drain_allow = false;
1469                                }
1470
1471                                if (!isolate_lru_page(head)) {
1472                                        list_add_tail(&head->lru, &cma_page_list);
1473                                        mod_node_page_state(page_pgdat(head),
1474                                                            NR_ISOLATED_ANON +
1475                                                            page_is_file_cache(head),
1476                                                            hpage_nr_pages(head));
1477                                }
1478                        }
1479                }
1480
1481                i += step;
1482        }
1483
1484        if (!list_empty(&cma_page_list)) {
1485                /*
1486                 * drop the above get_user_pages reference.
1487                 */
1488                for (i = 0; i < nr_pages; i++)
1489                        put_page(pages[i]);
1490
1491                if (migrate_pages(&cma_page_list, new_non_cma_page,
1492                                  NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1493                        /*
1494                         * some of the pages failed migration. Do get_user_pages
1495                         * without migration.
1496                         */
1497                        migrate_allow = false;
1498
1499                        if (!list_empty(&cma_page_list))
1500                                putback_movable_pages(&cma_page_list);
1501                }
1502                /*
1503                 * We did migrate all the pages, Try to get the page references
1504                 * again migrating any new CMA pages which we failed to isolate
1505                 * earlier.
1506                 */
1507                nr_pages = __get_user_pages_locked(tsk, mm, start, nr_pages,
1508                                                   pages, vmas, NULL,
1509                                                   gup_flags);
1510
1511                if ((nr_pages > 0) && migrate_allow) {
1512                        drain_allow = true;
1513                        goto check_again;
1514                }
1515        }
1516
1517        return nr_pages;
1518}
1519#else
1520static long check_and_migrate_cma_pages(struct task_struct *tsk,
1521                                        struct mm_struct *mm,
1522                                        unsigned long start,
1523                                        unsigned long nr_pages,
1524                                        struct page **pages,
1525                                        struct vm_area_struct **vmas,
1526                                        unsigned int gup_flags)
1527{
1528        return nr_pages;
1529}
1530#endif /* CONFIG_CMA */
1531
1532/*
1533 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1534 * allows us to process the FOLL_LONGTERM flag.
1535 */
1536static long __gup_longterm_locked(struct task_struct *tsk,
1537                                  struct mm_struct *mm,
1538                                  unsigned long start,
1539                                  unsigned long nr_pages,
1540                                  struct page **pages,
1541                                  struct vm_area_struct **vmas,
1542                                  unsigned int gup_flags)
1543{
1544        struct vm_area_struct **vmas_tmp = vmas;
1545        unsigned long flags = 0;
1546        long rc, i;
1547
1548        if (gup_flags & FOLL_LONGTERM) {
1549                if (!pages)
1550                        return -EINVAL;
1551
1552                if (!vmas_tmp) {
1553                        vmas_tmp = kcalloc(nr_pages,
1554                                           sizeof(struct vm_area_struct *),
1555                                           GFP_KERNEL);
1556                        if (!vmas_tmp)
1557                                return -ENOMEM;
1558                }
1559                flags = memalloc_nocma_save();
1560        }
1561
1562        rc = __get_user_pages_locked(tsk, mm, start, nr_pages, pages,
1563                                     vmas_tmp, NULL, gup_flags);
1564
1565        if (gup_flags & FOLL_LONGTERM) {
1566                memalloc_nocma_restore(flags);
1567                if (rc < 0)
1568                        goto out;
1569
1570                if (check_dax_vmas(vmas_tmp, rc)) {
1571                        for (i = 0; i < rc; i++)
1572                                put_page(pages[i]);
1573                        rc = -EOPNOTSUPP;
1574                        goto out;
1575                }
1576
1577                rc = check_and_migrate_cma_pages(tsk, mm, start, rc, pages,
1578                                                 vmas_tmp, gup_flags);
1579        }
1580
1581out:
1582        if (vmas_tmp != vmas)
1583                kfree(vmas_tmp);
1584        return rc;
1585}
1586#else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1587static __always_inline long __gup_longterm_locked(struct task_struct *tsk,
1588                                                  struct mm_struct *mm,
1589                                                  unsigned long start,
1590                                                  unsigned long nr_pages,
1591                                                  struct page **pages,
1592                                                  struct vm_area_struct **vmas,
1593                                                  unsigned int flags)
1594{
1595        return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1596                                       NULL, flags);
1597}
1598#endif /* CONFIG_FS_DAX || CONFIG_CMA */
1599
1600/*
1601 * This is the same as get_user_pages_remote(), just with a
1602 * less-flexible calling convention where we assume that the task
1603 * and mm being operated on are the current task's and don't allow
1604 * passing of a locked parameter.  We also obviously don't pass
1605 * FOLL_REMOTE in here.
1606 */
1607long get_user_pages(unsigned long start, unsigned long nr_pages,
1608                unsigned int gup_flags, struct page **pages,
1609                struct vm_area_struct **vmas)
1610{
1611        return __gup_longterm_locked(current, current->mm, start, nr_pages,
1612                                     pages, vmas, gup_flags | FOLL_TOUCH);
1613}
1614EXPORT_SYMBOL(get_user_pages);
1615
1616/*
1617 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1618 * paths better by using either get_user_pages_locked() or
1619 * get_user_pages_unlocked().
1620 *
1621 * get_user_pages_locked() is suitable to replace the form:
1622 *
1623 *      down_read(&mm->mmap_sem);
1624 *      do_something()
1625 *      get_user_pages(tsk, mm, ..., pages, NULL);
1626 *      up_read(&mm->mmap_sem);
1627 *
1628 *  to:
1629 *
1630 *      int locked = 1;
1631 *      down_read(&mm->mmap_sem);
1632 *      do_something()
1633 *      get_user_pages_locked(tsk, mm, ..., pages, &locked);
1634 *      if (locked)
1635 *          up_read(&mm->mmap_sem);
1636 */
1637long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1638                           unsigned int gup_flags, struct page **pages,
1639                           int *locked)
1640{
1641        /*
1642         * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1643         * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1644         * vmas.  As there are no users of this flag in this call we simply
1645         * disallow this option for now.
1646         */
1647        if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1648                return -EINVAL;
1649
1650        return __get_user_pages_locked(current, current->mm, start, nr_pages,
1651                                       pages, NULL, locked,
1652                                       gup_flags | FOLL_TOUCH);
1653}
1654EXPORT_SYMBOL(get_user_pages_locked);
1655
1656/*
1657 * get_user_pages_unlocked() is suitable to replace the form:
1658 *
1659 *      down_read(&mm->mmap_sem);
1660 *      get_user_pages(tsk, mm, ..., pages, NULL);
1661 *      up_read(&mm->mmap_sem);
1662 *
1663 *  with:
1664 *
1665 *      get_user_pages_unlocked(tsk, mm, ..., pages);
1666 *
1667 * It is functionally equivalent to get_user_pages_fast so
1668 * get_user_pages_fast should be used instead if specific gup_flags
1669 * (e.g. FOLL_FORCE) are not required.
1670 */
1671long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1672                             struct page **pages, unsigned int gup_flags)
1673{
1674        struct mm_struct *mm = current->mm;
1675        int locked = 1;
1676        long ret;
1677
1678        /*
1679         * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1680         * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1681         * vmas.  As there are no users of this flag in this call we simply
1682         * disallow this option for now.
1683         */
1684        if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1685                return -EINVAL;
1686
1687        down_read(&mm->mmap_sem);
1688        ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1689                                      &locked, gup_flags | FOLL_TOUCH);
1690        if (locked)
1691                up_read(&mm->mmap_sem);
1692        return ret;
1693}
1694EXPORT_SYMBOL(get_user_pages_unlocked);
1695
1696/*
1697 * Fast GUP
1698 *
1699 * get_user_pages_fast attempts to pin user pages by walking the page
1700 * tables directly and avoids taking locks. Thus the walker needs to be
1701 * protected from page table pages being freed from under it, and should
1702 * block any THP splits.
1703 *
1704 * One way to achieve this is to have the walker disable interrupts, and
1705 * rely on IPIs from the TLB flushing code blocking before the page table
1706 * pages are freed. This is unsuitable for architectures that do not need
1707 * to broadcast an IPI when invalidating TLBs.
1708 *
1709 * Another way to achieve this is to batch up page table containing pages
1710 * belonging to more than one mm_user, then rcu_sched a callback to free those
1711 * pages. Disabling interrupts will allow the fast_gup walker to both block
1712 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1713 * (which is a relatively rare event). The code below adopts this strategy.
1714 *
1715 * Before activating this code, please be aware that the following assumptions
1716 * are currently made:
1717 *
1718 *  *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1719 *  free pages containing page tables or TLB flushing requires IPI broadcast.
1720 *
1721 *  *) ptes can be read atomically by the architecture.
1722 *
1723 *  *) access_ok is sufficient to validate userspace address ranges.
1724 *
1725 * The last two assumptions can be relaxed by the addition of helper functions.
1726 *
1727 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1728 */
1729#ifdef CONFIG_HAVE_FAST_GUP
1730#ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
1731/*
1732 * WARNING: only to be used in the get_user_pages_fast() implementation.
1733 *
1734 * With get_user_pages_fast(), we walk down the pagetables without taking any
1735 * locks.  For this we would like to load the pointers atomically, but sometimes
1736 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE).  What
1737 * we do have is the guarantee that a PTE will only either go from not present
1738 * to present, or present to not present or both -- it will not switch to a
1739 * completely different present page without a TLB flush in between; something
1740 * that we are blocking by holding interrupts off.
1741 *
1742 * Setting ptes from not present to present goes:
1743 *
1744 *   ptep->pte_high = h;
1745 *   smp_wmb();
1746 *   ptep->pte_low = l;
1747 *
1748 * And present to not present goes:
1749 *
1750 *   ptep->pte_low = 0;
1751 *   smp_wmb();
1752 *   ptep->pte_high = 0;
1753 *
1754 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
1755 * We load pte_high *after* loading pte_low, which ensures we don't see an older
1756 * value of pte_high.  *Then* we recheck pte_low, which ensures that we haven't
1757 * picked up a changed pte high. We might have gotten rubbish values from
1758 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
1759 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
1760 * operates on present ptes we're safe.
1761 */
1762static inline pte_t gup_get_pte(pte_t *ptep)
1763{
1764        pte_t pte;
1765
1766        do {
1767                pte.pte_low = ptep->pte_low;
1768                smp_rmb();
1769                pte.pte_high = ptep->pte_high;
1770                smp_rmb();
1771        } while (unlikely(pte.pte_low != ptep->pte_low));
1772
1773        return pte;
1774}
1775#else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1776/*
1777 * We require that the PTE can be read atomically.
1778 */
1779static inline pte_t gup_get_pte(pte_t *ptep)
1780{
1781        return READ_ONCE(*ptep);
1782}
1783#endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1784
1785static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
1786                                            struct page **pages)
1787{
1788        while ((*nr) - nr_start) {
1789                struct page *page = pages[--(*nr)];
1790
1791                ClearPageReferenced(page);
1792                put_page(page);
1793        }
1794}
1795
1796/*
1797 * Return the compund head page with ref appropriately incremented,
1798 * or NULL if that failed.
1799 */
1800static inline struct page *try_get_compound_head(struct page *page, int refs)
1801{
1802        struct page *head = compound_head(page);
1803        if (WARN_ON_ONCE(page_ref_count(head) < 0))
1804                return NULL;
1805        if (unlikely(!page_cache_add_speculative(head, refs)))
1806                return NULL;
1807        return head;
1808}
1809
1810#ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1811static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1812                         unsigned int flags, struct page **pages, int *nr)
1813{
1814        struct dev_pagemap *pgmap = NULL;
1815        int nr_start = *nr, ret = 0;
1816        pte_t *ptep, *ptem;
1817
1818        ptem = ptep = pte_offset_map(&pmd, addr);
1819        do {
1820                pte_t pte = gup_get_pte(ptep);
1821                struct page *head, *page;
1822
1823                /*
1824                 * Similar to the PMD case below, NUMA hinting must take slow
1825                 * path using the pte_protnone check.
1826                 */
1827                if (pte_protnone(pte))
1828                        goto pte_unmap;
1829
1830                if (!pte_access_permitted(pte, flags & FOLL_WRITE))
1831                        goto pte_unmap;
1832
1833                if (pte_devmap(pte)) {
1834                        if (unlikely(flags & FOLL_LONGTERM))
1835                                goto pte_unmap;
1836
1837                        pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1838                        if (unlikely(!pgmap)) {
1839                                undo_dev_pagemap(nr, nr_start, pages);
1840                                goto pte_unmap;
1841                        }
1842                } else if (pte_special(pte))
1843                        goto pte_unmap;
1844
1845                VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1846                page = pte_page(pte);
1847
1848                head = try_get_compound_head(page, 1);
1849                if (!head)
1850                        goto pte_unmap;
1851
1852                if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1853                        put_page(head);
1854                        goto pte_unmap;
1855                }
1856
1857                VM_BUG_ON_PAGE(compound_head(page) != head, page);
1858
1859                SetPageReferenced(page);
1860                pages[*nr] = page;
1861                (*nr)++;
1862
1863        } while (ptep++, addr += PAGE_SIZE, addr != end);
1864
1865        ret = 1;
1866
1867pte_unmap:
1868        if (pgmap)
1869                put_dev_pagemap(pgmap);
1870        pte_unmap(ptem);
1871        return ret;
1872}
1873#else
1874
1875/*
1876 * If we can't determine whether or not a pte is special, then fail immediately
1877 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1878 * to be special.
1879 *
1880 * For a futex to be placed on a THP tail page, get_futex_key requires a
1881 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1882 * useful to have gup_huge_pmd even if we can't operate on ptes.
1883 */
1884static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1885                         unsigned int flags, struct page **pages, int *nr)
1886{
1887        return 0;
1888}
1889#endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1890
1891#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1892static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1893                unsigned long end, struct page **pages, int *nr)
1894{
1895        int nr_start = *nr;
1896        struct dev_pagemap *pgmap = NULL;
1897
1898        do {
1899                struct page *page = pfn_to_page(pfn);
1900
1901                pgmap = get_dev_pagemap(pfn, pgmap);
1902                if (unlikely(!pgmap)) {
1903                        undo_dev_pagemap(nr, nr_start, pages);
1904                        return 0;
1905                }
1906                SetPageReferenced(page);
1907                pages[*nr] = page;
1908                get_page(page);
1909                (*nr)++;
1910                pfn++;
1911        } while (addr += PAGE_SIZE, addr != end);
1912
1913        if (pgmap)
1914                put_dev_pagemap(pgmap);
1915        return 1;
1916}
1917
1918static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1919                unsigned long end, struct page **pages, int *nr)
1920{
1921        unsigned long fault_pfn;
1922        int nr_start = *nr;
1923
1924        fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1925        if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1926                return 0;
1927
1928        if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1929                undo_dev_pagemap(nr, nr_start, pages);
1930                return 0;
1931        }
1932        return 1;
1933}
1934
1935static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1936                unsigned long end, struct page **pages, int *nr)
1937{
1938        unsigned long fault_pfn;
1939        int nr_start = *nr;
1940
1941        fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1942        if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1943                return 0;
1944
1945        if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1946                undo_dev_pagemap(nr, nr_start, pages);
1947                return 0;
1948        }
1949        return 1;
1950}
1951#else
1952static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1953                unsigned long end, struct page **pages, int *nr)
1954{
1955        BUILD_BUG();
1956        return 0;
1957}
1958
1959static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1960                unsigned long end, struct page **pages, int *nr)
1961{
1962        BUILD_BUG();
1963        return 0;
1964}
1965#endif
1966
1967#ifdef CONFIG_ARCH_HAS_HUGEPD
1968static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
1969                                      unsigned long sz)
1970{
1971        unsigned long __boundary = (addr + sz) & ~(sz-1);
1972        return (__boundary - 1 < end - 1) ? __boundary : end;
1973}
1974
1975static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1976                       unsigned long end, unsigned int flags,
1977                       struct page **pages, int *nr)
1978{
1979        unsigned long pte_end;
1980        struct page *head, *page;
1981        pte_t pte;
1982        int refs;
1983
1984        pte_end = (addr + sz) & ~(sz-1);
1985        if (pte_end < end)
1986                end = pte_end;
1987
1988        pte = READ_ONCE(*ptep);
1989
1990        if (!pte_access_permitted(pte, flags & FOLL_WRITE))
1991                return 0;
1992
1993        /* hugepages are never "special" */
1994        VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1995
1996        refs = 0;
1997        head = pte_page(pte);
1998
1999        page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2000        do {
2001                VM_BUG_ON(compound_head(page) != head);
2002                pages[*nr] = page;
2003                (*nr)++;
2004                page++;
2005                refs++;
2006        } while (addr += PAGE_SIZE, addr != end);
2007
2008        head = try_get_compound_head(head, refs);
2009        if (!head) {
2010                *nr -= refs;
2011                return 0;
2012        }
2013
2014        if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2015                /* Could be optimized better */
2016                *nr -= refs;
2017                while (refs--)
2018                        put_page(head);
2019                return 0;
2020        }
2021
2022        SetPageReferenced(head);
2023        return 1;
2024}
2025
2026static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2027                unsigned int pdshift, unsigned long end, unsigned int flags,
2028                struct page **pages, int *nr)
2029{
2030        pte_t *ptep;
2031        unsigned long sz = 1UL << hugepd_shift(hugepd);
2032        unsigned long next;
2033
2034        ptep = hugepte_offset(hugepd, addr, pdshift);
2035        do {
2036                next = hugepte_addr_end(addr, end, sz);
2037                if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2038                        return 0;
2039        } while (ptep++, addr = next, addr != end);
2040
2041        return 1;
2042}
2043#else
2044static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2045                unsigned int pdshift, unsigned long end, unsigned int flags,
2046                struct page **pages, int *nr)
2047{
2048        return 0;
2049}
2050#endif /* CONFIG_ARCH_HAS_HUGEPD */
2051
2052static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2053                        unsigned long end, unsigned int flags,
2054                        struct page **pages, int *nr)
2055{
2056        struct page *head, *page;
2057        int refs;
2058
2059        if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2060                return 0;
2061
2062        if (pmd_devmap(orig)) {
2063                if (unlikely(flags & FOLL_LONGTERM))
2064                        return 0;
2065                return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
2066        }
2067
2068        refs = 0;
2069        page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2070        do {
2071                pages[*nr] = page;
2072                (*nr)++;
2073                page++;
2074                refs++;
2075        } while (addr += PAGE_SIZE, addr != end);
2076
2077        head = try_get_compound_head(pmd_page(orig), refs);
2078        if (!head) {
2079                *nr -= refs;
2080                return 0;
2081        }
2082
2083        if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2084                *nr -= refs;
2085                while (refs--)
2086                        put_page(head);
2087                return 0;
2088        }
2089
2090        SetPageReferenced(head);
2091        return 1;
2092}
2093
2094static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2095                unsigned long end, unsigned int flags, struct page **pages, int *nr)
2096{
2097        struct page *head, *page;
2098        int refs;
2099
2100        if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2101                return 0;
2102
2103        if (pud_devmap(orig)) {
2104                if (unlikely(flags & FOLL_LONGTERM))
2105                        return 0;
2106                return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
2107        }
2108
2109        refs = 0;
2110        page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2111        do {
2112                pages[*nr] = page;
2113                (*nr)++;
2114                page++;
2115                refs++;
2116        } while (addr += PAGE_SIZE, addr != end);
2117
2118        head = try_get_compound_head(pud_page(orig), refs);
2119        if (!head) {
2120                *nr -= refs;
2121                return 0;
2122        }
2123
2124        if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2125                *nr -= refs;
2126                while (refs--)
2127                        put_page(head);
2128                return 0;
2129        }
2130
2131        SetPageReferenced(head);
2132        return 1;
2133}
2134
2135static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2136                        unsigned long end, unsigned int flags,
2137                        struct page **pages, int *nr)
2138{
2139        int refs;
2140        struct page *head, *page;
2141
2142        if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2143                return 0;
2144
2145        BUILD_BUG_ON(pgd_devmap(orig));
2146        refs = 0;
2147        page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2148        do {
2149                pages[*nr] = page;
2150                (*nr)++;
2151                page++;
2152                refs++;
2153        } while (addr += PAGE_SIZE, addr != end);
2154
2155        head = try_get_compound_head(pgd_page(orig), refs);
2156        if (!head) {
2157                *nr -= refs;
2158                return 0;
2159        }
2160
2161        if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2162                *nr -= refs;
2163                while (refs--)
2164                        put_page(head);
2165                return 0;
2166        }
2167
2168        SetPageReferenced(head);
2169        return 1;
2170}
2171
2172static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
2173                unsigned int flags, struct page **pages, int *nr)
2174{
2175        unsigned long next;
2176        pmd_t *pmdp;
2177
2178        pmdp = pmd_offset(&pud, addr);
2179        do {
2180                pmd_t pmd = READ_ONCE(*pmdp);
2181
2182                next = pmd_addr_end(addr, end);
2183                if (!pmd_present(pmd))
2184                        return 0;
2185
2186                if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2187                             pmd_devmap(pmd))) {
2188                        /*
2189                         * NUMA hinting faults need to be handled in the GUP
2190                         * slowpath for accounting purposes and so that they
2191                         * can be serialised against THP migration.
2192                         */
2193                        if (pmd_protnone(pmd))
2194                                return 0;
2195
2196                        if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2197                                pages, nr))
2198                                return 0;
2199
2200                } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2201                        /*
2202                         * architecture have different format for hugetlbfs
2203                         * pmd format and THP pmd format
2204                         */
2205                        if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2206                                         PMD_SHIFT, next, flags, pages, nr))
2207                                return 0;
2208                } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2209                        return 0;
2210        } while (pmdp++, addr = next, addr != end);
2211
2212        return 1;
2213}
2214
2215static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
2216                         unsigned int flags, struct page **pages, int *nr)
2217{
2218        unsigned long next;
2219        pud_t *pudp;
2220
2221        pudp = pud_offset(&p4d, addr);
2222        do {
2223                pud_t pud = READ_ONCE(*pudp);
2224
2225                next = pud_addr_end(addr, end);
2226                if (pud_none(pud))
2227                        return 0;
2228                if (unlikely(pud_huge(pud))) {
2229                        if (!gup_huge_pud(pud, pudp, addr, next, flags,
2230                                          pages, nr))
2231                                return 0;
2232                } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2233                        if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2234                                         PUD_SHIFT, next, flags, pages, nr))
2235                                return 0;
2236                } else if (!gup_pmd_range(pud, addr, next, flags, pages, nr))
2237                        return 0;
2238        } while (pudp++, addr = next, addr != end);
2239
2240        return 1;
2241}
2242
2243static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
2244                         unsigned int flags, struct page **pages, int *nr)
2245{
2246        unsigned long next;
2247        p4d_t *p4dp;
2248
2249        p4dp = p4d_offset(&pgd, addr);
2250        do {
2251                p4d_t p4d = READ_ONCE(*p4dp);
2252
2253                next = p4d_addr_end(addr, end);
2254                if (p4d_none(p4d))
2255                        return 0;
2256                BUILD_BUG_ON(p4d_huge(p4d));
2257                if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2258                        if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2259                                         P4D_SHIFT, next, flags, pages, nr))
2260                                return 0;
2261                } else if (!gup_pud_range(p4d, addr, next, flags, pages, nr))
2262                        return 0;
2263        } while (p4dp++, addr = next, addr != end);
2264
2265        return 1;
2266}
2267
2268static void gup_pgd_range(unsigned long addr, unsigned long end,
2269                unsigned int flags, struct page **pages, int *nr)
2270{
2271        unsigned long next;
2272        pgd_t *pgdp;
2273
2274        pgdp = pgd_offset(current->mm, addr);
2275        do {
2276                pgd_t pgd = READ_ONCE(*pgdp);
2277
2278                next = pgd_addr_end(addr, end);
2279                if (pgd_none(pgd))
2280                        return;
2281                if (unlikely(pgd_huge(pgd))) {
2282                        if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2283                                          pages, nr))
2284                                return;
2285                } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2286                        if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2287                                         PGDIR_SHIFT, next, flags, pages, nr))
2288                                return;
2289                } else if (!gup_p4d_range(pgd, addr, next, flags, pages, nr))
2290                        return;
2291        } while (pgdp++, addr = next, addr != end);
2292}
2293#else
2294static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2295                unsigned int flags, struct page **pages, int *nr)
2296{
2297}
2298#endif /* CONFIG_HAVE_FAST_GUP */
2299
2300#ifndef gup_fast_permitted
2301/*
2302 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2303 * we need to fall back to the slow version:
2304 */
2305static bool gup_fast_permitted(unsigned long start, unsigned long end)
2306{
2307        return true;
2308}
2309#endif
2310
2311/*
2312 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2313 * the regular GUP.
2314 * Note a difference with get_user_pages_fast: this always returns the
2315 * number of pages pinned, 0 if no pages were pinned.
2316 *
2317 * If the architecture does not support this function, simply return with no
2318 * pages pinned.
2319 */
2320int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
2321                          struct page **pages)
2322{
2323        unsigned long len, end;
2324        unsigned long flags;
2325        int nr = 0;
2326
2327        start = untagged_addr(start) & PAGE_MASK;
2328        len = (unsigned long) nr_pages << PAGE_SHIFT;
2329        end = start + len;
2330
2331        if (end <= start)
2332                return 0;
2333        if (unlikely(!access_ok((void __user *)start, len)))
2334                return 0;
2335
2336        /*
2337         * Disable interrupts.  We use the nested form as we can already have
2338         * interrupts disabled by get_futex_key.
2339         *
2340         * With interrupts disabled, we block page table pages from being
2341         * freed from under us. See struct mmu_table_batch comments in
2342         * include/asm-generic/tlb.h for more details.
2343         *
2344         * We do not adopt an rcu_read_lock(.) here as we also want to
2345         * block IPIs that come from THPs splitting.
2346         */
2347
2348        if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2349            gup_fast_permitted(start, end)) {
2350                local_irq_save(flags);
2351                gup_pgd_range(start, end, write ? FOLL_WRITE : 0, pages, &nr);
2352                local_irq_restore(flags);
2353        }
2354
2355        return nr;
2356}
2357EXPORT_SYMBOL_GPL(__get_user_pages_fast);
2358
2359static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2360                                   unsigned int gup_flags, struct page **pages)
2361{
2362        int ret;
2363
2364        /*
2365         * FIXME: FOLL_LONGTERM does not work with
2366         * get_user_pages_unlocked() (see comments in that function)
2367         */
2368        if (gup_flags & FOLL_LONGTERM) {
2369                down_read(&current->mm->mmap_sem);
2370                ret = __gup_longterm_locked(current, current->mm,
2371                                            start, nr_pages,
2372                                            pages, NULL, gup_flags);
2373                up_read(&current->mm->mmap_sem);
2374        } else {
2375                ret = get_user_pages_unlocked(start, nr_pages,
2376                                              pages, gup_flags);
2377        }
2378
2379        return ret;
2380}
2381
2382/**
2383 * get_user_pages_fast() - pin user pages in memory
2384 * @start:      starting user address
2385 * @nr_pages:   number of pages from start to pin
2386 * @gup_flags:  flags modifying pin behaviour
2387 * @pages:      array that receives pointers to the pages pinned.
2388 *              Should be at least nr_pages long.
2389 *
2390 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2391 * If not successful, it will fall back to taking the lock and
2392 * calling get_user_pages().
2393 *
2394 * Returns number of pages pinned. This may be fewer than the number
2395 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2396 * were pinned, returns -errno.
2397 */
2398int get_user_pages_fast(unsigned long start, int nr_pages,
2399                        unsigned int gup_flags, struct page **pages)
2400{
2401        unsigned long addr, len, end;
2402        int nr = 0, ret = 0;
2403
2404        if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM)))
2405                return -EINVAL;
2406
2407        start = untagged_addr(start) & PAGE_MASK;
2408        addr = start;
2409        len = (unsigned long) nr_pages << PAGE_SHIFT;
2410        end = start + len;
2411
2412        if (end <= start)
2413                return 0;
2414        if (unlikely(!access_ok((void __user *)start, len)))
2415                return -EFAULT;
2416
2417        if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2418            gup_fast_permitted(start, end)) {
2419                local_irq_disable();
2420                gup_pgd_range(addr, end, gup_flags, pages, &nr);
2421                local_irq_enable();
2422                ret = nr;
2423        }
2424
2425        if (nr < nr_pages) {
2426                /* Try to get the remaining pages with get_user_pages */
2427                start += nr << PAGE_SHIFT;
2428                pages += nr;
2429
2430                ret = __gup_longterm_unlocked(start, nr_pages - nr,
2431                                              gup_flags, pages);
2432
2433                /* Have to be a bit careful with return values */
2434                if (nr > 0) {
2435                        if (ret < 0)
2436                                ret = nr;
2437                        else
2438                                ret += nr;
2439                }
2440        }
2441
2442        return ret;
2443}
2444EXPORT_SYMBOL_GPL(get_user_pages_fast);
2445