mm, dax: fix DAX deadlocks
authorRoss Zwisler <ross.zwisler@linux.intel.com>
Thu, 15 Oct 2015 22:28:32 +0000 (15:28 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 16 Oct 2015 18:42:28 +0000 (11:42 -0700)
The following two locking commits in the DAX code:

commit 843172978bb9 ("dax: fix race between simultaneous faults")
commit 46c043ede471 ("mm: take i_mmap_lock in unmap_mapping_range() for DAX")

introduced a number of deadlocks and other issues which need to be fixed
for the v4.3 kernel.  The list of issues in DAX after these commits
(some newly introduced by the commits, some preexisting) can be found
here:

  https://lkml.org/lkml/2015/9/25/602 (Subject: "Re: [PATCH] dax: fix deadlock in __dax_fault").

This undoes most of the changes introduced by those two commits,
essentially returning us to the DAX locking scheme that was used in
v4.2.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Dan Williams <dan.j.williams@intel.com>
Tested-by: Dave Chinner <dchinner@redhat.com>
Cc: Jan Kara <jack@suse.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Matthew Wilcox <matthew.r.wilcox@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/dax.c
mm/memory.c

index bcfb14bfc1e49eb36d507c53ccd1c3f4f23135f9..a86d3cc2b38941b0e39f23be84e4986d8852ae42 100644 (file)
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -285,6 +285,7 @@ static int copy_user_bh(struct page *to, struct buffer_head *bh,
 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
                        struct vm_area_struct *vma, struct vm_fault *vmf)
 {
+       struct address_space *mapping = inode->i_mapping;
        sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
        unsigned long vaddr = (unsigned long)vmf->virtual_address;
        void __pmem *addr;
@@ -292,6 +293,8 @@ static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
        pgoff_t size;
        int error;
 
+       i_mmap_lock_read(mapping);
+
        /*
         * Check truncate didn't happen while we were allocating a block.
         * If it did, this block may or may not be still allocated to the
@@ -321,6 +324,8 @@ static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
        error = vm_insert_mixed(vma, vaddr, pfn);
 
  out:
+       i_mmap_unlock_read(mapping);
+
        return error;
 }
 
@@ -382,17 +387,15 @@ int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
                         * from a read fault and we've raced with a truncate
                         */
                        error = -EIO;
-                       goto unlock;
+                       goto unlock_page;
                }
-       } else {
-               i_mmap_lock_write(mapping);
        }
 
        error = get_block(inode, block, &bh, 0);
        if (!error && (bh.b_size < PAGE_SIZE))
                error = -EIO;           /* fs corruption? */
        if (error)
-               goto unlock;
+               goto unlock_page;
 
        if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
                if (vmf->flags & FAULT_FLAG_WRITE) {
@@ -403,9 +406,8 @@ int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
                        if (!error && (bh.b_size < PAGE_SIZE))
                                error = -EIO;
                        if (error)
-                               goto unlock;
+                               goto unlock_page;
                } else {
-                       i_mmap_unlock_write(mapping);
                        return dax_load_hole(mapping, page, vmf);
                }
        }
@@ -417,15 +419,17 @@ int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
                else
                        clear_user_highpage(new_page, vaddr);
                if (error)
-                       goto unlock;
+                       goto unlock_page;
                vmf->page = page;
                if (!page) {
+                       i_mmap_lock_read(mapping);
                        /* Check we didn't race with truncate */
                        size = (i_size_read(inode) + PAGE_SIZE - 1) >>
                                                                PAGE_SHIFT;
                        if (vmf->pgoff >= size) {
+                               i_mmap_unlock_read(mapping);
                                error = -EIO;
-                               goto unlock;
+                               goto out;
                        }
                }
                return VM_FAULT_LOCKED;
@@ -461,8 +465,6 @@ int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
                        WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
        }
 
-       if (!page)
-               i_mmap_unlock_write(mapping);
  out:
        if (error == -ENOMEM)
                return VM_FAULT_OOM | major;
@@ -471,14 +473,11 @@ int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
                return VM_FAULT_SIGBUS | major;
        return VM_FAULT_NOPAGE | major;
 
- unlock:
+ unlock_page:
        if (page) {
                unlock_page(page);
                page_cache_release(page);
-       } else {
-               i_mmap_unlock_write(mapping);
        }
-
        goto out;
 }
 EXPORT_SYMBOL(__dax_fault);
@@ -556,10 +555,10 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
        block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
 
        bh.b_size = PMD_SIZE;
-       i_mmap_lock_write(mapping);
        length = get_block(inode, block, &bh, write);
        if (length)
                return VM_FAULT_SIGBUS;
+       i_mmap_lock_read(mapping);
 
        /*
         * If the filesystem isn't willing to tell us the length of a hole,
@@ -569,36 +568,14 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
        if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE)
                goto fallback;
 
-       sector = bh.b_blocknr << (blkbits - 9);
-
-       if (buffer_unwritten(&bh) || buffer_new(&bh)) {
-               int i;
-
-               length = bdev_direct_access(bh.b_bdev, sector, &kaddr, &pfn,
-                                               bh.b_size);
-               if (length < 0) {
-                       result = VM_FAULT_SIGBUS;
-                       goto out;
-               }
-               if ((length < PMD_SIZE) || (pfn & PG_PMD_COLOUR))
-                       goto fallback;
-
-               for (i = 0; i < PTRS_PER_PMD; i++)
-                       clear_pmem(kaddr + i * PAGE_SIZE, PAGE_SIZE);
-               wmb_pmem();
-               count_vm_event(PGMAJFAULT);
-               mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
-               result |= VM_FAULT_MAJOR;
-       }
-
        /*
         * If we allocated new storage, make sure no process has any
         * zero pages covering this hole
         */
        if (buffer_new(&bh)) {
-               i_mmap_unlock_write(mapping);
+               i_mmap_unlock_read(mapping);
                unmap_mapping_range(mapping, pgoff << PAGE_SHIFT, PMD_SIZE, 0);
-               i_mmap_lock_write(mapping);
+               i_mmap_lock_read(mapping);
        }
 
        /*
@@ -635,6 +612,7 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
                result = VM_FAULT_NOPAGE;
                spin_unlock(ptl);
        } else {
+               sector = bh.b_blocknr << (blkbits - 9);
                length = bdev_direct_access(bh.b_bdev, sector, &kaddr, &pfn,
                                                bh.b_size);
                if (length < 0) {
@@ -644,15 +622,25 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
                if ((length < PMD_SIZE) || (pfn & PG_PMD_COLOUR))
                        goto fallback;
 
+               if (buffer_unwritten(&bh) || buffer_new(&bh)) {
+                       int i;
+                       for (i = 0; i < PTRS_PER_PMD; i++)
+                               clear_pmem(kaddr + i * PAGE_SIZE, PAGE_SIZE);
+                       wmb_pmem();
+                       count_vm_event(PGMAJFAULT);
+                       mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
+                       result |= VM_FAULT_MAJOR;
+               }
+
                result |= vmf_insert_pfn_pmd(vma, address, pmd, pfn, write);
        }
 
  out:
+       i_mmap_unlock_read(mapping);
+
        if (buffer_unwritten(&bh))
                complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
 
-       i_mmap_unlock_write(mapping);
-
        return result;
 
  fallback:
index 9cb27470fee991cb874676bb0cbc0f694b5e1d36..deb679c31f2ab897cafebf72643aec4f66233308 100644 (file)
@@ -2426,6 +2426,8 @@ void unmap_mapping_range(struct address_space *mapping,
        if (details.last_index < details.first_index)
                details.last_index = ULONG_MAX;
 
+
+       /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
        i_mmap_lock_write(mapping);
        if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
                unmap_mapping_range_tree(&mapping->i_mmap, &details);