releases/5.0.10/coredump-fix-race-condition-between-mmget_not_zero-get_task_mm-and-core-dumping.patch

   1 From 04f5866e41fb70690e28397487d8bd8eea7d712a Mon Sep 17 00:00:00 2001
   2 From: Andrea Arcangeli <aarcange@redhat.com>
   3 Date: Thu, 18 Apr 2019 17:50:52 -0700
   4 Subject: coredump: fix race condition between mmget_not_zero()/get_task_mm() and core dumping
   5
   6 From: Andrea Arcangeli <aarcange@redhat.com>
   7
   8 commit 04f5866e41fb70690e28397487d8bd8eea7d712a upstream.
   9
  10 The core dumping code has always run without holding the mmap_sem for
  11 writing, despite that is the only way to ensure that the entire vma
  12 layout will not change from under it.  Only using some signal
  13 serialization on the processes belonging to the mm is not nearly enough.
  14 This was pointed out earlier.  For example in Hugh's post from Jul 2017:
  15
  16   https://lkml.kernel.org/r/alpine.LSU.2.11.1707191716030.2055@eggly.anvils
  17
  18   "Not strictly relevant here, but a related note: I was very surprised
  19    to discover, only quite recently, how handle_mm_fault() may be called
  20    without down_read(mmap_sem) - when core dumping. That seems a
  21    misguided optimization to me, which would also be nice to correct"
  22
  23 In particular because the growsdown and growsup can move the
  24 vm_start/vm_end the various loops the core dump does around the vma will
  25 not be consistent if page faults can happen concurrently.
  26
  27 Pretty much all users calling mmget_not_zero()/get_task_mm() and then
  28 taking the mmap_sem had the potential to introduce unexpected side
  29 effects in the core dumping code.
  30
  31 Adding mmap_sem for writing around the ->core_dump invocation is a
  32 viable long term fix, but it requires removing all copy user and page
  33 faults and to replace them with get_dump_page() for all binary formats
  34 which is not suitable as a short term fix.
  35
  36 For the time being this solution manually covers the places that can
  37 confuse the core dump either by altering the vma layout or the vma flags
  38 while it runs.  Once ->core_dump runs under mmap_sem for writing the
  39 function mmget_still_valid() can be dropped.
  40
  41 Allowing mmap_sem protected sections to run in parallel with the
  42 coredump provides some minor parallelism advantage to the swapoff code
  43 (which seems to be safe enough by never mangling any vma field and can
  44 keep doing swapins in parallel to the core dumping) and to some other
  45 corner case.
  46
  47 In order to facilitate the backporting I added "Fixes: 86039bd3b4e6"
  48 however the side effect of this same race condition in /proc/pid/mem
  49 should be reproducible since before 2.6.12-rc2 so I couldn't add any
  50 other "Fixes:" because there's no hash beyond the git genesis commit.
  51
  52 Because find_extend_vma() is the only location outside of the process
  53 context that could modify the "mm" structures under mmap_sem for
  54 reading, by adding the mmget_still_valid() check to it, all other cases
  55 that take the mmap_sem for reading don't need the new check after
  56 mmget_not_zero()/get_task_mm().  The expand_stack() in page fault
  57 context also doesn't need the new check, because all tasks under core
  58 dumping are frozen.
  59
  60 Link: http://lkml.kernel.org/r/20190325224949.11068-1-aarcange@redhat.com
  61 Fixes: 86039bd3b4e6 ("userfaultfd: add new syscall to provide memory externalization")
  62 Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
  63 Reported-by: Jann Horn <jannh@google.com>
  64 Suggested-by: Oleg Nesterov <oleg@redhat.com>
  65 Acked-by: Peter Xu <peterx@redhat.com>
  66 Reviewed-by: Mike Rapoport <rppt@linux.ibm.com>
  67 Reviewed-by: Oleg Nesterov <oleg@redhat.com>
  68 Reviewed-by: Jann Horn <jannh@google.com>
  69 Acked-by: Jason Gunthorpe <jgg@mellanox.com>
  70 Acked-by: Michal Hocko <mhocko@suse.com>
  71 Cc: <stable@vger.kernel.org>
  72 Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
  73 Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
  74 Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
  75
  76 ---
  77  drivers/infiniband/core/uverbs_main.c |    3 +++
  78  fs/proc/task_mmu.c                    |   18 ++++++++++++++++++
  79  fs/userfaultfd.c                      |    9 +++++++++
  80  include/linux/sched/mm.h              |   21 +++++++++++++++++++++
  81  mm/mmap.c                             |    7 ++++++-
  82  5 files changed, 57 insertions(+), 1 deletion(-)
  83
  84 --- a/drivers/infiniband/core/uverbs_main.c
  85 +++ b/drivers/infiniband/core/uverbs_main.c
  86 @@ -992,6 +992,8 @@ void uverbs_user_mmap_disassociate(struc
  87                  * will only be one mm, so no big deal.
  88                  */
  89                 down_write(&mm->mmap_sem);
  90 +               if (!mmget_still_valid(mm))
  91 +                       goto skip_mm;
  92                 mutex_lock(&ufile->umap_lock);
  93                 list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
  94                                           list) {
  95 @@ -1006,6 +1008,7 @@ void uverbs_user_mmap_disassociate(struc
  96                         vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
  97                 }
  98                 mutex_unlock(&ufile->umap_lock);
  99 +       skip_mm:
 100                 up_write(&mm->mmap_sem);
 101                 mmput(mm);
 102         }
 103 --- a/fs/proc/task_mmu.c
 104 +++ b/fs/proc/task_mmu.c
 105 @@ -1141,6 +1141,24 @@ static ssize_t clear_refs_write(struct f
 106                                         count = -EINTR;
 107                                         goto out_mm;
 108                                 }
 109 +                               /*
 110 +                                * Avoid to modify vma->vm_flags
 111 +                                * without locked ops while the
 112 +                                * coredump reads the vm_flags.
 113 +                                */
 114 +                               if (!mmget_still_valid(mm)) {
 115 +                                       /*
 116 +                                        * Silently return "count"
 117 +                                        * like if get_task_mm()
 118 +                                        * failed. FIXME: should this
 119 +                                        * function have returned
 120 +                                        * -ESRCH if get_task_mm()
 121 +                                        * failed like if
 122 +                                        * get_proc_task() fails?
 123 +                                        */
 124 +                                       up_write(&mm->mmap_sem);
 125 +                                       goto out_mm;
 126 +                               }
 127                                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
 128                                         vma->vm_flags &= ~VM_SOFTDIRTY;
 129                                         vma_set_page_prot(vma);
 130 --- a/fs/userfaultfd.c
 131 +++ b/fs/userfaultfd.c
 132 @@ -629,6 +629,8 @@ static void userfaultfd_event_wait_compl
 133
 134                 /* the various vma->vm_userfaultfd_ctx still points to it */
 135                 down_write(&mm->mmap_sem);
 136 +               /* no task can run (and in turn coredump) yet */
 137 +               VM_WARN_ON(!mmget_still_valid(mm));
 138                 for (vma = mm->mmap; vma; vma = vma->vm_next)
 139                         if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
 140                                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 141 @@ -883,6 +885,8 @@ static int userfaultfd_release(struct in
 142          * taking the mmap_sem for writing.
 143          */
 144         down_write(&mm->mmap_sem);
 145 +       if (!mmget_still_valid(mm))
 146 +               goto skip_mm;
 147         prev = NULL;
 148         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 149                 cond_resched();
 150 @@ -905,6 +909,7 @@ static int userfaultfd_release(struct in
 151                 vma->vm_flags = new_flags;
 152                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 153         }
 154 +skip_mm:
 155         up_write(&mm->mmap_sem);
 156         mmput(mm);
 157  wakeup:
 158 @@ -1333,6 +1338,8 @@ static int userfaultfd_register(struct u
 159                 goto out;
 160
 161         down_write(&mm->mmap_sem);
 162 +       if (!mmget_still_valid(mm))
 163 +               goto out_unlock;
 164         vma = find_vma_prev(mm, start, &prev);
 165         if (!vma)
 166                 goto out_unlock;
 167 @@ -1520,6 +1527,8 @@ static int userfaultfd_unregister(struct
 168                 goto out;
 169
 170         down_write(&mm->mmap_sem);
 171 +       if (!mmget_still_valid(mm))
 172 +               goto out_unlock;
 173         vma = find_vma_prev(mm, start, &prev);
 174         if (!vma)
 175                 goto out_unlock;
 176 --- a/include/linux/sched/mm.h
 177 +++ b/include/linux/sched/mm.h
 178 @@ -49,6 +49,27 @@ static inline void mmdrop(struct mm_stru
 179                 __mmdrop(mm);
 180  }
 181
 182 +/*
 183 + * This has to be called after a get_task_mm()/mmget_not_zero()
 184 + * followed by taking the mmap_sem for writing before modifying the
 185 + * vmas or anything the coredump pretends not to change from under it.
 186 + *
 187 + * NOTE: find_extend_vma() called from GUP context is the only place
 188 + * that can modify the "mm" (notably the vm_start/end) under mmap_sem
 189 + * for reading and outside the context of the process, so it is also
 190 + * the only case that holds the mmap_sem for reading that must call
 191 + * this function. Generally if the mmap_sem is hold for reading
 192 + * there's no need of this check after get_task_mm()/mmget_not_zero().
 193 + *
 194 + * This function can be obsoleted and the check can be removed, after
 195 + * the coredump code will hold the mmap_sem for writing before
 196 + * invoking the ->core_dump methods.
 197 + */
 198 +static inline bool mmget_still_valid(struct mm_struct *mm)
 199 +{
 200 +       return likely(!mm->core_state);
 201 +}
 202 +
 203  /**
 204   * mmget() - Pin the address space associated with a &struct mm_struct.
 205   * @mm: The address space to pin.
 206 --- a/mm/mmap.c
 207 +++ b/mm/mmap.c
 208 @@ -45,6 +45,7 @@
 209  #include <linux/moduleparam.h>
 210  #include <linux/pkeys.h>
 211  #include <linux/oom.h>
 212 +#include <linux/sched/mm.h>
 213
 214  #include <linux/uaccess.h>
 215  #include <asm/cacheflush.h>
 216 @@ -2526,7 +2527,8 @@ find_extend_vma(struct mm_struct *mm, un
 217         vma = find_vma_prev(mm, addr, &prev);
 218         if (vma && (vma->vm_start <= addr))
 219                 return vma;
 220 -       if (!prev || expand_stack(prev, addr))
 221 +       /* don't alter vm_end if the coredump is running */
 222 +       if (!prev || !mmget_still_valid(mm) || expand_stack(prev, addr))
 223                 return NULL;
 224         if (prev->vm_flags & VM_LOCKED)
 225                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
 226 @@ -2552,6 +2554,9 @@ find_extend_vma(struct mm_struct *mm, un
 227                 return vma;
 228         if (!(vma->vm_flags & VM_GROWSDOWN))
 229                 return NULL;
 230 +       /* don't alter vm_start if the coredump is running */
 231 +       if (!mmget_still_valid(mm))
 232 +               return NULL;
 233         start = vma->vm_start;
 234         if (expand_stack(vma, addr))
 235                 return NULL;