Merge tag 'dax-locking-for-4.7' of git://git.kernel.org/pub/scm/linux/kernel/git...
[sfrench/cifs-2.6.git] / mm / memory.c
index a1b93d9e444981445e1b8e5f50ff914492156b8f..15322b73636b4272d6c84af6742a0822822301db 100644 (file)
@@ -63,6 +63,7 @@
 #include <linux/dma-debug.h>
 #include <linux/debugfs.h>
 #include <linux/userfaultfd_k.h>
+#include <linux/dax.h>
 
 #include <asm/io.h>
 #include <asm/mmu_context.h>
@@ -2492,8 +2493,6 @@ 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);
@@ -2825,7 +2824,8 @@ oom:
  */
 static int __do_fault(struct vm_area_struct *vma, unsigned long address,
                        pgoff_t pgoff, unsigned int flags,
-                       struct page *cow_page, struct page **page)
+                       struct page *cow_page, struct page **page,
+                       void **entry)
 {
        struct vm_fault vmf;
        int ret;
@@ -2840,8 +2840,10 @@ static int __do_fault(struct vm_area_struct *vma, unsigned long address,
        ret = vma->vm_ops->fault(vma, &vmf);
        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
                return ret;
-       if (!vmf.page)
-               goto out;
+       if (ret & VM_FAULT_DAX_LOCKED) {
+               *entry = vmf.entry;
+               return ret;
+       }
 
        if (unlikely(PageHWPoison(vmf.page))) {
                if (ret & VM_FAULT_LOCKED)
@@ -2855,7 +2857,6 @@ static int __do_fault(struct vm_area_struct *vma, unsigned long address,
        else
                VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
 
- out:
        *page = vmf.page;
        return ret;
 }
@@ -3048,7 +3049,7 @@ static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
                pte_unmap_unlock(pte, ptl);
        }
 
-       ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
+       ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page, NULL);
        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
                return ret;
 
@@ -3071,6 +3072,7 @@ static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
                pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
 {
        struct page *fault_page, *new_page;
+       void *fault_entry;
        struct mem_cgroup *memcg;
        spinlock_t *ptl;
        pte_t *pte;
@@ -3088,26 +3090,24 @@ static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
                return VM_FAULT_OOM;
        }
 
-       ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
+       ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page,
+                        &fault_entry);
        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
                goto uncharge_out;
 
-       if (fault_page)
+       if (!(ret & VM_FAULT_DAX_LOCKED))
                copy_user_highpage(new_page, fault_page, address, vma);
        __SetPageUptodate(new_page);
 
        pte = pte_offset_map_lock(mm, pmd, address, &ptl);
        if (unlikely(!pte_same(*pte, orig_pte))) {
                pte_unmap_unlock(pte, ptl);
-               if (fault_page) {
+               if (!(ret & VM_FAULT_DAX_LOCKED)) {
                        unlock_page(fault_page);
                        put_page(fault_page);
                } else {
-                       /*
-                        * The fault handler has no page to lock, so it holds
-                        * i_mmap_lock for read to protect against truncate.
-                        */
-                       i_mmap_unlock_read(vma->vm_file->f_mapping);
+                       dax_unlock_mapping_entry(vma->vm_file->f_mapping,
+                                                pgoff);
                }
                goto uncharge_out;
        }
@@ -3115,15 +3115,11 @@ static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        mem_cgroup_commit_charge(new_page, memcg, false, false);
        lru_cache_add_active_or_unevictable(new_page, vma);
        pte_unmap_unlock(pte, ptl);
-       if (fault_page) {
+       if (!(ret & VM_FAULT_DAX_LOCKED)) {
                unlock_page(fault_page);
                put_page(fault_page);
        } else {
-               /*
-                * The fault handler has no page to lock, so it holds
-                * i_mmap_lock for read to protect against truncate.
-                */
-               i_mmap_unlock_read(vma->vm_file->f_mapping);
+               dax_unlock_mapping_entry(vma->vm_file->f_mapping, pgoff);
        }
        return ret;
 uncharge_out:
@@ -3143,7 +3139,7 @@ static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        int dirtied = 0;
        int ret, tmp;
 
-       ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
+       ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page, NULL);
        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
                return ret;