]> www.infradead.org Git - users/jedix/linux-maple.git/commitdiff
mm/mmap: Use advanced maple tree API for mmap_region()
authorLiam R. Howlett <Liam.Howlett@Oracle.com>
Tue, 10 Nov 2020 18:37:40 +0000 (13:37 -0500)
committerLiam R. Howlett <Liam.Howlett@oracle.com>
Fri, 26 Nov 2021 19:50:29 +0000 (14:50 -0500)
Changing mmap_region() to use the maple tree state and the advanced
maple tree interface allows for a lot less tree walking.

This change removes the last caller of munmap_vma_range(), so drop this
unused function.

Add vma_expand() to expand a VMA if possible by doing the necessary
hugepage check, uprobe_munmap of files, dcache flush, modifications then
undoing the detaches, etc.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
mm/mmap.c

index c06c5b850e1e5e36ae642c18ce5a51f5bceb44f2..b0b7e327bf8bf031fa1ba0a2d6ff2ef24aa1c4a0 100644 (file)
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -496,29 +496,6 @@ static inline struct vm_area_struct *__vma_next(struct mm_struct *mm,
        return vma->vm_next;
 }
 
-/*
- * munmap_vma_range() - munmap VMAs that overlap a range.
- * @mm: The mm struct
- * @start: The start of the range.
- * @len: The length of the range.
- * @pprev: pointer to the pointer that will be set to previous vm_area_struct
- *
- * Find all the vm_area_struct that overlap from @start to
- * @end and munmap them.  Set @pprev to the previous vm_area_struct.
- *
- * Returns: -ENOMEM on munmap failure or 0 on success.
- */
-static inline int
-munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
-                struct vm_area_struct **pprev, struct list_head *uf)
-{
-       // Needs optimization.
-       while (range_has_overlap(mm, start, start + len, pprev))
-               if (do_munmap(mm, start, len, uf))
-                       return -ENOMEM;
-       return 0;
-}
-
 static unsigned long count_vma_pages_range(struct mm_struct *mm,
                unsigned long addr, unsigned long end)
 {
@@ -619,6 +596,101 @@ static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
        mm->map_count++;
 }
 
+/*
+ * vma_expand - Expand an existing VMA
+ * @mas: The maple state
+ * @vma: The vma to expand
+ * @start: The start of the vma
+ * @end: The exclusive end of the vma
+ */
+inline int vma_expand(struct ma_state *mas, struct vm_area_struct *vma,
+                     unsigned long start, unsigned long end, pgoff_t pgoff,
+                     struct vm_area_struct *next)
+{
+       struct mm_struct *mm = vma->vm_mm;
+       struct address_space *mapping = NULL;
+       struct rb_root_cached *root = NULL;
+       struct anon_vma *anon_vma = vma->anon_vma;
+       struct file *file = vma->vm_file;
+       bool remove_next = false;
+       int error;
+
+       if (next && (vma != next) && (end == next->vm_end)) {
+               remove_next = true;
+               if (next->anon_vma && !vma->anon_vma) {
+                       vma->anon_vma = next->anon_vma;
+                       error = anon_vma_clone(vma, next);
+                       if (error)
+                               return error;
+               }
+       }
+
+       vma_adjust_trans_huge(vma, start, end, 0);
+
+       if (file) {
+               mapping = file->f_mapping;
+               root = &mapping->i_mmap;
+               uprobe_munmap(vma, vma->vm_start, vma->vm_end);
+               i_mmap_lock_write(mapping);
+               flush_dcache_mmap_lock(mapping);
+               vma_interval_tree_remove(vma, root);
+       } else if (anon_vma) {
+               anon_vma_lock_write(anon_vma);
+               anon_vma_interval_tree_pre_update_vma(vma);
+       }
+
+       vma->vm_start = start;
+       vma->vm_end = end;
+       vma->vm_pgoff = pgoff;
+       /* Note: mas must be pointing to the expanding VMA */
+       vma_mas_store(vma, mas);
+
+       if (file) {
+               vma_interval_tree_insert(vma, root);
+               flush_dcache_mmap_unlock(mapping);
+       }
+
+       /* Expanding over the next vma */
+       if (remove_next) {
+               /* Remove from mm linked list - also updates highest_vm_end */
+               __vma_unlink_list(mm, next);
+
+               /* Kill the cache */
+               vmacache_invalidate(mm);
+
+               if (file)
+                       __remove_shared_vm_struct(next, file, mapping);
+
+       } else if (!next) {
+               mm->highest_vm_end = vm_end_gap(vma);
+       }
+
+       if (anon_vma) {
+               anon_vma_interval_tree_post_update_vma(vma);
+               anon_vma_unlock_write(anon_vma);
+       }
+
+       if (file) {
+               i_mmap_unlock_write(mapping);
+               uprobe_mmap(vma);
+       }
+
+       if (remove_next) {
+               if (file) {
+                       uprobe_munmap(next, next->vm_start, next->vm_end);
+                       fput(file);
+               }
+               if (next->anon_vma)
+                       anon_vma_merge(vma, next);
+               mm->map_count--;
+               mpol_put(vma_policy(next));
+               vm_area_free(next);
+       }
+
+       validate_mm(mm);
+       return 0;
+}
+
 /*
  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
  * is already present in an i_mmap tree without adjusting the tree.
@@ -1598,9 +1670,15 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
                struct list_head *uf)
 {
        struct mm_struct *mm = current->mm;
-       struct vm_area_struct *vma, *prev, *merge;
-       int error;
+       struct vm_area_struct *vma = NULL;
+       struct vm_area_struct *prev, *next;
+       pgoff_t pglen = len >> PAGE_SHIFT;
        unsigned long charged = 0;
+       unsigned long end = addr + len;
+       unsigned long merge_start = addr, merge_end = end;
+       pgoff_t vm_pgoff;
+       int error;
+       MA_STATE(mas, &mm->mm_mt, addr, end - 1);
 
        /* Check against address space limit. */
        if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
@@ -1610,16 +1688,17 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
                 * MAP_FIXED may remove pages of mappings that intersects with
                 * requested mapping. Account for the pages it would unmap.
                 */
-               nr_pages = count_vma_pages_range(mm, addr, addr + len);
+               nr_pages = count_vma_pages_range(mm, addr, end);
 
                if (!may_expand_vm(mm, vm_flags,
                                        (len >> PAGE_SHIFT) - nr_pages))
                        return -ENOMEM;
        }
 
-       /* Clear old maps, set up prev and uf */
-       if (munmap_vma_range(mm, addr, len, &prev, uf))
+       /* Unmap any existing mapping in the area */
+       if (do_munmap(mm, addr, len, uf))
                return -ENOMEM;
+
        /*
         * Private writable mapping: check memory availability
         */
@@ -1630,14 +1709,41 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
                vm_flags |= VM_ACCOUNT;
        }
 
-       /*
-        * Can we just expand an old mapping?
-        */
-       vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
-                       NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
-       if (vma)
-               goto out;
+       next = mas_next(&mas, ULONG_MAX);
+       prev = mas_prev(&mas, 0);
+       if (vm_flags & VM_SPECIAL)
+               goto cannot_expand;
+
+       /* Attempt to expand an old mapping */
+       /* Check next */
+       if (next && next->vm_start == end && vma_policy(next) &&
+           can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
+                                NULL_VM_UFFD_CTX)) {
+               merge_end = next->vm_end;
+               vma = next;
+               vm_pgoff = next->vm_pgoff - pglen;
+       }
 
+       /* Check prev */
+       if (prev && prev->vm_end == addr && !vma_policy(prev) &&
+           can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
+                               NULL_VM_UFFD_CTX)) {
+               merge_start = prev->vm_start;
+               vma = prev;
+               vm_pgoff = prev->vm_pgoff;
+       }
+
+
+       /* Actually expand, if possible */
+       if (vma &&
+           !vma_expand(&mas, vma, merge_start, merge_end, vm_pgoff, next)) {
+               khugepaged_enter_vma_merge(prev, vm_flags);
+               goto expanded;
+       }
+
+       mas.index = addr;
+       mas.last = end - 1;
+cannot_expand:
        /*
         * Determine the object being mapped and call the appropriate
         * specific mapper. the address has already been validated, but
@@ -1650,7 +1756,7 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
        }
 
        vma->vm_start = addr;
-       vma->vm_end = addr + len;
+       vma->vm_end = end;
        vma->vm_flags = vm_flags;
        vma->vm_page_prot = vm_get_page_prot(vm_flags);
        vma->vm_pgoff = pgoff;
@@ -1671,8 +1777,6 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
                 *
                 * Answer: Yes, several device drivers can do it in their
                 *         f_op->mmap method. -DaveM
-                * Bug: If addr is changed, prev, rb_link, rb_parent should
-                *      be updated for vma_link()
                 */
                WARN_ON_ONCE(addr != vma->vm_start);
 
@@ -1681,23 +1785,31 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
                /* If vm_flags changed after call_mmap(), we should try merge vma again
                 * as we may succeed this time.
                 */
-               if (unlikely(vm_flags != vma->vm_flags && prev)) {
-                       merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
-                               NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
-                       if (merge) {
+               if (unlikely(vm_flags != vma->vm_flags && prev &&
+                            prev->vm_end == addr && !vma_policy(prev) &&
+                            can_vma_merge_after(prev, vm_flags, NULL, file,
+                                                pgoff, NULL_VM_UFFD_CTX))) {
+                       merge_start = prev->vm_start;
+                       vm_pgoff = prev->vm_pgoff;
+                       if (!vma_expand(&mas, prev, merge_start, merge_end,
+                                       vm_pgoff, next)) {
                                /* ->mmap() can change vma->vm_file and fput the original file. So
                                 * fput the vma->vm_file here or we would add an extra fput for file
                                 * and cause general protection fault ultimately.
                                 */
                                fput(vma->vm_file);
                                vm_area_free(vma);
-                               vma = merge;
-                               /* Update vm_flags to pick up the change. */
+                               vma = prev;
+                               /* Update vm_flags and possible addr to pick up the change. We don't
+                                * warn here if addr changed as the vma is not linked by vma_link().
+                                */
+                               addr = vma->vm_start;
                                vm_flags = vma->vm_flags;
                                goto unmap_writable;
                        }
                }
 
+               mas_set(&mas, addr);
                vm_flags = vma->vm_flags;
        } else if (vm_flags & VM_SHARED) {
                error = shmem_zero_setup(vma);
@@ -1716,20 +1828,35 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
                        goto free_vma;
        }
 
-       vma_link(mm, vma, prev);
+       if (vma->vm_file)
+               i_mmap_lock_write(vma->vm_file->f_mapping);
+
+       vma_mas_store(vma, &mas);
+       __vma_link_list(mm, vma, prev);
+       mm->map_count++;
+       if (vma->vm_file) {
+               if (vma->vm_flags & VM_SHARED)
+                       mapping_allow_writable(vma->vm_file->f_mapping);
+
+               flush_dcache_mmap_lock(vma->vm_file->f_mapping);
+               vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
+               flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
+               i_mmap_unlock_write(vma->vm_file->f_mapping);
+       }
+
        /* Once vma denies write, undo our temporary denial count */
 unmap_writable:
        if (file && vm_flags & VM_SHARED)
                mapping_unmap_writable(file->f_mapping);
        file = vma->vm_file;
-out:
+expanded:
        perf_event_mmap(vma);
 
        vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
        if (vm_flags & VM_LOCKED) {
                if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
-                                       is_vm_hugetlb_page(vma) ||
-                                       vma == get_gate_vma(current->mm))
+                   is_vm_hugetlb_page(vma) ||
+                   vma == get_gate_vma(current->mm))
                        vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
                else
                        mm->locked_vm += (len >> PAGE_SHIFT);
@@ -2557,10 +2684,6 @@ int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
        prev = vma->vm_prev;
        /* we have start < vma->vm_end  */
 
-       /* if it doesn't overlap, we have nothing.. */
-       if (vma->vm_start >= end)
-               return 0;
-
        /*
         * If we need to split any vma, do it now to save pain later.
         *