]> git.itanic.dy.fi Git - linux-stable/commitdiff
mm/madvise: make MADV_POPULATE_(READ|WRITE) handle VM_FAULT_RETRY properly
authorDavid Hildenbrand <david@redhat.com>
Thu, 14 Mar 2024 16:12:59 +0000 (17:12 +0100)
committerAndrew Morton <akpm@linux-foundation.org>
Tue, 16 Apr 2024 22:39:48 +0000 (15:39 -0700)
Darrick reports that in some cases where pread() would fail with -EIO and
mmap()+access would generate a SIGBUS signal, MADV_POPULATE_READ /
MADV_POPULATE_WRITE will keep retrying forever and not fail with -EFAULT.

While the madvise() call can be interrupted by a signal, this is not the
desired behavior.  MADV_POPULATE_READ / MADV_POPULATE_WRITE should behave
like page faults in that case: fail and not retry forever.

A reproducer can be found at [1].

The reason is that __get_user_pages(), as called by
faultin_vma_page_range(), will not handle VM_FAULT_RETRY in a proper way:
it will simply return 0 when VM_FAULT_RETRY happened, making
madvise_populate()->faultin_vma_page_range() retry again and again, never
setting FOLL_TRIED->FAULT_FLAG_TRIED for __get_user_pages().

__get_user_pages_locked() does what we want, but duplicating that logic in
faultin_vma_page_range() feels wrong.

So let's use __get_user_pages_locked() instead, that will detect
VM_FAULT_RETRY and set FOLL_TRIED when retrying, making the fault handler
return VM_FAULT_SIGBUS (VM_FAULT_ERROR) at some point, propagating -EFAULT
from faultin_page() to __get_user_pages(), all the way to
madvise_populate().

But, there is an issue: __get_user_pages_locked() will end up re-taking
the MM lock and then __get_user_pages() will do another VMA lookup.  In
the meantime, the VMA layout could have changed and we'd fail with
different error codes than we'd want to.

As __get_user_pages() will currently do a new VMA lookup either way, let
it do the VMA handling in a different way, controlled by a new
FOLL_MADV_POPULATE flag, effectively moving these checks from
madvise_populate() + faultin_page_range() in there.

With this change, Darricks reproducer properly fails with -EFAULT, as
documented for MADV_POPULATE_READ / MADV_POPULATE_WRITE.

[1] https://lore.kernel.org/all/20240313171936.GN1927156@frogsfrogsfrogs/

Link: https://lkml.kernel.org/r/20240314161300.382526-1-david@redhat.com
Link: https://lkml.kernel.org/r/20240314161300.382526-2-david@redhat.com
Fixes: 4ca9b3859dac ("mm/madvise: introduce MADV_POPULATE_(READ|WRITE) to prefault page tables")
Signed-off-by: David Hildenbrand <david@redhat.com>
Reported-by: Darrick J. Wong <djwong@kernel.org>
Closes: https://lore.kernel.org/all/20240311223815.GW1927156@frogsfrogsfrogs/
Cc: Darrick J. Wong <djwong@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
mm/gup.c
mm/internal.h
mm/madvise.c

index af8edadc05d1b87200050bd34ff3d58ec52abd52..1611e73b1121b1b9031356ccac6b765c60327ba1 100644 (file)
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -1206,6 +1206,22 @@ static long __get_user_pages(struct mm_struct *mm,
 
                /* first iteration or cross vma bound */
                if (!vma || start >= vma->vm_end) {
+                       /*
+                        * MADV_POPULATE_(READ|WRITE) wants to handle VMA
+                        * lookups+error reporting differently.
+                        */
+                       if (gup_flags & FOLL_MADV_POPULATE) {
+                               vma = vma_lookup(mm, start);
+                               if (!vma) {
+                                       ret = -ENOMEM;
+                                       goto out;
+                               }
+                               if (check_vma_flags(vma, gup_flags)) {
+                                       ret = -EINVAL;
+                                       goto out;
+                               }
+                               goto retry;
+                       }
                        vma = gup_vma_lookup(mm, start);
                        if (!vma && in_gate_area(mm, start)) {
                                ret = get_gate_page(mm, start & PAGE_MASK,
@@ -1685,35 +1701,35 @@ long populate_vma_page_range(struct vm_area_struct *vma,
 }
 
 /*
- * faultin_vma_page_range() - populate (prefault) page tables inside the
- *                           given VMA range readable/writable
+ * faultin_page_range() - populate (prefault) page tables inside the
+ *                       given range readable/writable
  *
  * This takes care of mlocking the pages, too, if VM_LOCKED is set.
  *
- * @vma: target vma
+ * @mm: the mm to populate page tables in
  * @start: start address
  * @end: end address
  * @write: whether to prefault readable or writable
  * @locked: whether the mmap_lock is still held
  *
- * Returns either number of processed pages in the vma, or a negative error
- * code on error (see __get_user_pages()).
+ * Returns either number of processed pages in the MM, or a negative error
+ * code on error (see __get_user_pages()). Note that this function reports
+ * errors related to VMAs, such as incompatible mappings, as expected by
+ * MADV_POPULATE_(READ|WRITE).
  *
- * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and
- * covered by the VMA. If it's released, *@locked will be set to 0.
+ * The range must be page-aligned.
+ *
+ * mm->mmap_lock must be held. If it's released, *@locked will be set to 0.
  */
-long faultin_vma_page_range(struct vm_area_struct *vma, unsigned long start,
-                           unsigned long end, bool write, int *locked)
+long faultin_page_range(struct mm_struct *mm, unsigned long start,
+                       unsigned long end, bool write, int *locked)
 {
-       struct mm_struct *mm = vma->vm_mm;
        unsigned long nr_pages = (end - start) / PAGE_SIZE;
        int gup_flags;
        long ret;
 
        VM_BUG_ON(!PAGE_ALIGNED(start));
        VM_BUG_ON(!PAGE_ALIGNED(end));
-       VM_BUG_ON_VMA(start < vma->vm_start, vma);
-       VM_BUG_ON_VMA(end > vma->vm_end, vma);
        mmap_assert_locked(mm);
 
        /*
@@ -1725,19 +1741,13 @@ long faultin_vma_page_range(struct vm_area_struct *vma, unsigned long start,
         *                a poisoned page.
         * !FOLL_FORCE: Require proper access permissions.
         */
-       gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE;
+       gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE |
+                   FOLL_MADV_POPULATE;
        if (write)
                gup_flags |= FOLL_WRITE;
 
-       /*
-        * We want to report -EINVAL instead of -EFAULT for any permission
-        * problems or incompatible mappings.
-        */
-       if (check_vma_flags(vma, gup_flags))
-               return -EINVAL;
-
-       ret = __get_user_pages(mm, start, nr_pages, gup_flags,
-                              NULL, locked);
+       ret = __get_user_pages_locked(mm, start, nr_pages, NULL, locked,
+                                     gup_flags);
        lru_add_drain();
        return ret;
 }
index 7e486f2c502cee245991e2468a0655228a81aef5..07ad2675a88b4798b140d3af3a303a20519cddc3 100644 (file)
@@ -686,9 +686,8 @@ struct anon_vma *folio_anon_vma(struct folio *folio);
 void unmap_mapping_folio(struct folio *folio);
 extern long populate_vma_page_range(struct vm_area_struct *vma,
                unsigned long start, unsigned long end, int *locked);
-extern long faultin_vma_page_range(struct vm_area_struct *vma,
-                                  unsigned long start, unsigned long end,
-                                  bool write, int *locked);
+extern long faultin_page_range(struct mm_struct *mm, unsigned long start,
+               unsigned long end, bool write, int *locked);
 extern bool mlock_future_ok(struct mm_struct *mm, unsigned long flags,
                               unsigned long bytes);
 
@@ -1127,10 +1126,13 @@ enum {
        FOLL_FAST_ONLY = 1 << 20,
        /* allow unlocking the mmap lock */
        FOLL_UNLOCKABLE = 1 << 21,
+       /* VMA lookup+checks compatible with MADV_POPULATE_(READ|WRITE) */
+       FOLL_MADV_POPULATE = 1 << 22,
 };
 
 #define INTERNAL_GUP_FLAGS (FOLL_TOUCH | FOLL_TRIED | FOLL_REMOTE | FOLL_PIN | \
-                           FOLL_FAST_ONLY | FOLL_UNLOCKABLE)
+                           FOLL_FAST_ONLY | FOLL_UNLOCKABLE | \
+                           FOLL_MADV_POPULATE)
 
 /*
  * Indicates for which pages that are write-protected in the page table,
index 44a498c94158c882c624eac2e29a5f07d854e322..1a073fcc4c0c021496667619a20f6ee1afaef7c0 100644 (file)
@@ -908,27 +908,14 @@ static long madvise_populate(struct vm_area_struct *vma,
 {
        const bool write = behavior == MADV_POPULATE_WRITE;
        struct mm_struct *mm = vma->vm_mm;
-       unsigned long tmp_end;
        int locked = 1;
        long pages;
 
        *prev = vma;
 
        while (start < end) {
-               /*
-                * We might have temporarily dropped the lock. For example,
-                * our VMA might have been split.
-                */
-               if (!vma || start >= vma->vm_end) {
-                       vma = vma_lookup(mm, start);
-                       if (!vma)
-                               return -ENOMEM;
-               }
-
-               tmp_end = min_t(unsigned long, end, vma->vm_end);
                /* Populate (prefault) page tables readable/writable. */
-               pages = faultin_vma_page_range(vma, start, tmp_end, write,
-                                              &locked);
+               pages = faultin_page_range(mm, start, end, write, &locked);
                if (!locked) {
                        mmap_read_lock(mm);
                        locked = 1;
@@ -949,7 +936,7 @@ static long madvise_populate(struct vm_area_struct *vma,
                                pr_warn_once("%s: unhandled return value: %ld\n",
                                             __func__, pages);
                                fallthrough;
-                       case -ENOMEM:
+                       case -ENOMEM: /* No VMA or out of memory. */
                                return -ENOMEM;
                        }
                }