hugetlbfs: revert "Use i_mmap_rwsem to fix page fault/truncate race"
authorMike Kravetz <mike.kravetz@oracle.com>
Tue, 8 Jan 2019 23:23:32 +0000 (15:23 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 9 Jan 2019 01:15:11 +0000 (17:15 -0800)
This reverts c86aa7bbfd5568ba8a82d3635d8f7b8a8e06fe54

The reverted commit caused ABBA deadlocks when file migration raced with
file eviction for specific hugetlbfs files.  This was discovered with a
modified version of the LTP move_pages12 test.

The purpose of the reverted patch was to close a long existing race
between hugetlbfs file truncation and page faults.  After more analysis
of the patch and impacted code, it was determined that i_mmap_rwsem can
not be used for all required synchronization.  Therefore, revert this
patch while working an another approach to the underlying issue.

Link: http://lkml.kernel.org/r/20190103235452.29335-1-mike.kravetz@oracle.com
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Reported-by: Jan Stancek <jstancek@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: "Aneesh Kumar K . V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Prakash Sangappa <prakash.sangappa@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/hugetlbfs/inode.c
mm/hugetlb.c

index a2fcea5f8225c1ec0228623b1f332ee02bba9c1c..32920a10100e23fc60f53cf36c882278ae972cee 100644 (file)
@@ -383,16 +383,17 @@ hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
  * truncation is indicated by end of range being LLONG_MAX
  *     In this case, we first scan the range and release found pages.
  *     After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
- *     maps and global counts.
+ *     maps and global counts.  Page faults can not race with truncation
+ *     in this routine.  hugetlb_no_page() prevents page faults in the
+ *     truncated range.  It checks i_size before allocation, and again after
+ *     with the page table lock for the page held.  The same lock must be
+ *     acquired to unmap a page.
  * hole punch is indicated if end is not LLONG_MAX
  *     In the hole punch case we scan the range and release found pages.
  *     Only when releasing a page is the associated region/reserv map
  *     deleted.  The region/reserv map for ranges without associated
- *     pages are not modified.
- *
- * Callers of this routine must hold the i_mmap_rwsem in write mode to prevent
- * races with page faults.
- *
+ *     pages are not modified.  Page faults can race with hole punch.
+ *     This is indicated if we find a mapped page.
  * Note: If the passed end of range value is beyond the end of file, but
  * not LLONG_MAX this routine still performs a hole punch operation.
  */
@@ -422,14 +423,32 @@ static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 
                for (i = 0; i < pagevec_count(&pvec); ++i) {
                        struct page *page = pvec.pages[i];
+                       u32 hash;
 
                        index = page->index;
+                       hash = hugetlb_fault_mutex_hash(h, current->mm,
+                                                       &pseudo_vma,
+                                                       mapping, index, 0);
+                       mutex_lock(&hugetlb_fault_mutex_table[hash]);
+
                        /*
-                        * A mapped page is impossible as callers should unmap
-                        * all references before calling.  And, i_mmap_rwsem
-                        * prevents the creation of additional mappings.
+                        * If page is mapped, it was faulted in after being
+                        * unmapped in caller.  Unmap (again) now after taking
+                        * the fault mutex.  The mutex will prevent faults
+                        * until we finish removing the page.
+                        *
+                        * This race can only happen in the hole punch case.
+                        * Getting here in a truncate operation is a bug.
                         */
-                       VM_BUG_ON(page_mapped(page));
+                       if (unlikely(page_mapped(page))) {
+                               BUG_ON(truncate_op);
+
+                               i_mmap_lock_write(mapping);
+                               hugetlb_vmdelete_list(&mapping->i_mmap,
+                                       index * pages_per_huge_page(h),
+                                       (index + 1) * pages_per_huge_page(h));
+                               i_mmap_unlock_write(mapping);
+                       }
 
                        lock_page(page);
                        /*
@@ -451,6 +470,7 @@ static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
                        }
 
                        unlock_page(page);
+                       mutex_unlock(&hugetlb_fault_mutex_table[hash]);
                }
                huge_pagevec_release(&pvec);
                cond_resched();
@@ -462,20 +482,9 @@ static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 
 static void hugetlbfs_evict_inode(struct inode *inode)
 {
-       struct address_space *mapping = inode->i_mapping;
        struct resv_map *resv_map;
 
-       /*
-        * The vfs layer guarantees that there are no other users of this
-        * inode.  Therefore, it would be safe to call remove_inode_hugepages
-        * without holding i_mmap_rwsem.  We acquire and hold here to be
-        * consistent with other callers.  Since there will be no contention
-        * on the semaphore, overhead is negligible.
-        */
-       i_mmap_lock_write(mapping);
        remove_inode_hugepages(inode, 0, LLONG_MAX);
-       i_mmap_unlock_write(mapping);
-
        resv_map = (struct resv_map *)inode->i_mapping->private_data;
        /* root inode doesn't have the resv_map, so we should check it */
        if (resv_map)
@@ -496,8 +505,8 @@ static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
        i_mmap_lock_write(mapping);
        if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
                hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
-       remove_inode_hugepages(inode, offset, LLONG_MAX);
        i_mmap_unlock_write(mapping);
+       remove_inode_hugepages(inode, offset, LLONG_MAX);
        return 0;
 }
 
@@ -531,8 +540,8 @@ static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
                        hugetlb_vmdelete_list(&mapping->i_mmap,
                                                hole_start >> PAGE_SHIFT,
                                                hole_end  >> PAGE_SHIFT);
-               remove_inode_hugepages(inode, hole_start, hole_end);
                i_mmap_unlock_write(mapping);
+               remove_inode_hugepages(inode, hole_start, hole_end);
                inode_unlock(inode);
        }
 
@@ -615,11 +624,7 @@ static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
                /* addr is the offset within the file (zero based) */
                addr = index * hpage_size;
 
-               /*
-                * fault mutex taken here, protects against fault path
-                * and hole punch.  inode_lock previously taken protects
-                * against truncation.
-                */
+               /* mutex taken here, fault path and hole punch */
                hash = hugetlb_fault_mutex_hash(h, mm, &pseudo_vma, mapping,
                                                index, addr);
                mutex_lock(&hugetlb_fault_mutex_table[hash]);
index 7450888109651f8974de4c04cab9a91ef7309932..aedc1b183cf935f4051103a270f779be8ac8263a 100644 (file)
@@ -3755,16 +3755,16 @@ static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
        }
 
        /*
-        * We can not race with truncation due to holding i_mmap_rwsem.
-        * Check once here for faults beyond end of file.
+        * Use page lock to guard against racing truncation
+        * before we get page_table_lock.
         */
-       size = i_size_read(mapping->host) >> huge_page_shift(h);
-       if (idx >= size)
-               goto out;
-
 retry:
        page = find_lock_page(mapping, idx);
        if (!page) {
+               size = i_size_read(mapping->host) >> huge_page_shift(h);
+               if (idx >= size)
+                       goto out;
+
                /*
                 * Check for page in userfault range
                 */
@@ -3854,6 +3854,9 @@ retry:
        }
 
        ptl = huge_pte_lock(h, mm, ptep);
+       size = i_size_read(mapping->host) >> huge_page_shift(h);
+       if (idx >= size)
+               goto backout;
 
        ret = 0;
        if (!huge_pte_none(huge_ptep_get(ptep)))
@@ -3956,10 +3959,8 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 
        /*
         * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
-        * until finished with ptep.  This serves two purposes:
-        * 1) It prevents huge_pmd_unshare from being called elsewhere
-        *    and making the ptep no longer valid.
-        * 2) It synchronizes us with file truncation.
+        * until finished with ptep.  This prevents huge_pmd_unshare from
+        * being called elsewhere and making the ptep no longer valid.
         *
         * ptep could have already be assigned via huge_pte_offset.  That
         * is OK, as huge_pte_alloc will return the same value unless