]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/commitdiff
dax: Call ->iomap_begin without entry lock during dax fault
authorJan Kara <jack@suse.cz>
Wed, 19 Oct 2016 12:34:31 +0000 (14:34 +0200)
committerDan Williams <dan.j.williams@intel.com>
Tue, 27 Dec 2016 04:29:25 +0000 (20:29 -0800)
Currently ->iomap_begin() handler is called with entry lock held. If the
filesystem held any locks between ->iomap_begin() and ->iomap_end()
(such as ext4 which will want to hold transaction open), this would cause
lock inversion with the iomap_apply() from standard IO path which first
calls ->iomap_begin() and only then calls ->actor() callback which grabs
entry locks for DAX (if it faults when copying from/to user provided
buffers).

Fix the problem by nesting grabbing of entry lock inside ->iomap_begin()
- ->iomap_end() pair.

Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
fs/dax.c

index bfec6f2ef613b9e9d11373f6d3a2cea0c2020808..5c74f60d0a5094dc0a27f27ae0acd41667414332 100644 (file)
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -1078,6 +1078,15 @@ dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
 }
 EXPORT_SYMBOL_GPL(dax_iomap_rw);
 
+static int dax_fault_return(int error)
+{
+       if (error == 0)
+               return VM_FAULT_NOPAGE;
+       if (error == -ENOMEM)
+               return VM_FAULT_OOM;
+       return VM_FAULT_SIGBUS;
+}
+
 /**
  * dax_iomap_fault - handle a page fault on a DAX file
  * @vma: The virtual memory area where the fault occurred
@@ -1110,12 +1119,6 @@ int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
        if (pos >= i_size_read(inode))
                return VM_FAULT_SIGBUS;
 
-       entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
-       if (IS_ERR(entry)) {
-               error = PTR_ERR(entry);
-               goto out;
-       }
-
        if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
                flags |= IOMAP_WRITE;
 
@@ -1126,9 +1129,15 @@ int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
         */
        error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
        if (error)
-               goto unlock_entry;
+               return dax_fault_return(error);
        if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
-               error = -EIO;           /* fs corruption? */
+               vmf_ret = dax_fault_return(-EIO);       /* fs corruption? */
+               goto finish_iomap;
+       }
+
+       entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
+       if (IS_ERR(entry)) {
+               vmf_ret = dax_fault_return(PTR_ERR(entry));
                goto finish_iomap;
        }
 
@@ -1151,13 +1160,13 @@ int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
                }
 
                if (error)
-                       goto finish_iomap;
+                       goto error_unlock_entry;
 
                __SetPageUptodate(vmf->cow_page);
                vmf_ret = finish_fault(vmf);
                if (!vmf_ret)
                        vmf_ret = VM_FAULT_DONE_COW;
-               goto finish_iomap;
+               goto unlock_entry;
        }
 
        switch (iomap.type) {
@@ -1169,12 +1178,15 @@ int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
                }
                error = dax_insert_mapping(mapping, iomap.bdev, sector,
                                PAGE_SIZE, &entry, vma, vmf);
+               /* -EBUSY is fine, somebody else faulted on the same PTE */
+               if (error == -EBUSY)
+                       error = 0;
                break;
        case IOMAP_UNWRITTEN:
        case IOMAP_HOLE:
                if (!(vmf->flags & FAULT_FLAG_WRITE)) {
                        vmf_ret = dax_load_hole(mapping, &entry, vmf);
-                       goto finish_iomap;
+                       goto unlock_entry;
                }
                /*FALLTHRU*/
        default:
@@ -1183,30 +1195,25 @@ int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
                break;
        }
 
- finish_iomap:
-       if (ops->iomap_end) {
-               if (error || (vmf_ret & VM_FAULT_ERROR)) {
-                       /* keep previous error */
-                       ops->iomap_end(inode, pos, PAGE_SIZE, 0, flags,
-                                       &iomap);
-               } else {
-                       error = ops->iomap_end(inode, pos, PAGE_SIZE,
-                                       PAGE_SIZE, flags, &iomap);
-               }
-       }
+ error_unlock_entry:
+       vmf_ret = dax_fault_return(error) | major;
  unlock_entry:
        put_locked_mapping_entry(mapping, vmf->pgoff, entry);
- out:
-       if (error == -ENOMEM)
-               return VM_FAULT_OOM | major;
-       /* -EBUSY is fine, somebody else faulted on the same PTE */
-       if (error < 0 && error != -EBUSY)
-               return VM_FAULT_SIGBUS | major;
-       if (vmf_ret) {
-               WARN_ON_ONCE(error); /* -EBUSY from ops->iomap_end? */
-               return vmf_ret;
+ finish_iomap:
+       if (ops->iomap_end) {
+               int copied = PAGE_SIZE;
+
+               if (vmf_ret & VM_FAULT_ERROR)
+                       copied = 0;
+               /*
+                * The fault is done by now and there's no way back (other
+                * thread may be already happily using PTE we have installed).
+                * Just ignore error from ->iomap_end since we cannot do much
+                * with it.
+                */
+               ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
        }
-       return VM_FAULT_NOPAGE | major;
+       return vmf_ret;
 }
 EXPORT_SYMBOL_GPL(dax_iomap_fault);
 
@@ -1330,16 +1337,6 @@ int dax_iomap_pmd_fault(struct vm_area_struct *vma, unsigned long address,
        if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
                goto fallback;
 
-       /*
-        * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
-        * PMD or a HZP entry.  If it can't (because a 4k page is already in
-        * the tree, for instance), it will return -EEXIST and we just fall
-        * back to 4k entries.
-        */
-       entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
-       if (IS_ERR(entry))
-               goto fallback;
-
        /*
         * Note that we don't use iomap_apply here.  We aren't doing I/O, only
         * setting up a mapping, so really we're using iomap_begin() as a way
@@ -1348,10 +1345,21 @@ int dax_iomap_pmd_fault(struct vm_area_struct *vma, unsigned long address,
        pos = (loff_t)pgoff << PAGE_SHIFT;
        error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
        if (error)
-               goto unlock_entry;
+               goto fallback;
+
        if (iomap.offset + iomap.length < pos + PMD_SIZE)
                goto finish_iomap;
 
+       /*
+        * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
+        * PMD or a HZP entry.  If it can't (because a 4k page is already in
+        * the tree, for instance), it will return -EEXIST and we just fall
+        * back to 4k entries.
+        */
+       entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
+       if (IS_ERR(entry))
+               goto finish_iomap;
+
        vmf.pgoff = pgoff;
        vmf.flags = flags;
        vmf.gfp_mask = mapping_gfp_mask(mapping) | __GFP_IO;
@@ -1364,7 +1372,7 @@ int dax_iomap_pmd_fault(struct vm_area_struct *vma, unsigned long address,
        case IOMAP_UNWRITTEN:
        case IOMAP_HOLE:
                if (WARN_ON_ONCE(write))
-                       goto finish_iomap;
+                       goto unlock_entry;
                result = dax_pmd_load_hole(vma, pmd, &vmf, address, &iomap,
                                &entry);
                break;
@@ -1373,20 +1381,23 @@ int dax_iomap_pmd_fault(struct vm_area_struct *vma, unsigned long address,
                break;
        }
 
+ unlock_entry:
+       put_locked_mapping_entry(mapping, pgoff, entry);
  finish_iomap:
        if (ops->iomap_end) {
-               if (result == VM_FAULT_FALLBACK) {
-                       ops->iomap_end(inode, pos, PMD_SIZE, 0, iomap_flags,
-                                       &iomap);
-               } else {
-                       error = ops->iomap_end(inode, pos, PMD_SIZE, PMD_SIZE,
-                                       iomap_flags, &iomap);
-                       if (error)
-                               result = VM_FAULT_FALLBACK;
-               }
+               int copied = PMD_SIZE;
+
+               if (result == VM_FAULT_FALLBACK)
+                       copied = 0;
+               /*
+                * The fault is done by now and there's no way back (other
+                * thread may be already happily using PMD we have installed).
+                * Just ignore error from ->iomap_end since we cannot do much
+                * with it.
+                */
+               ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
+                               &iomap);
        }
- unlock_entry:
-       put_locked_mapping_entry(mapping, pgoff, entry);
  fallback:
        if (result == VM_FAULT_FALLBACK) {
                split_huge_pmd(vma, pmd, address);