]> www.infradead.org Git - users/jedix/linux-maple.git/commitdiff
mm/mmap and friends: Remove rb tree. maple3_upstream howlett/maple/20200727
authorLiam R. Howlett <Liam.Howlett@Oracle.com>
Fri, 24 Jul 2020 19:06:25 +0000 (15:06 -0400)
committerLiam R. Howlett <Liam.Howlett@Oracle.com>
Sun, 26 Jul 2020 01:07:09 +0000 (21:07 -0400)
Remove the RB tree and start using the maple tree for vm_area_struct
tracking.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
drivers/firmware/efi/efi.c
include/linux/mm_types.h
kernel/fork.c
mm/init-mm.c
mm/mmap.c
mm/util.c

index 5114cae4ec9735dad8c0d809176957ce2ba676b2..8eaf97febbc87a446ecdf2a99b91fb1df74c992e 100644 (file)
@@ -51,7 +51,6 @@ static unsigned long __initdata mem_reserve = EFI_INVALID_TABLE_ADDR;
 static unsigned long __initdata rt_prop = EFI_INVALID_TABLE_ADDR;
 
 struct mm_struct efi_mm = {
-       .mm_rb                  = RB_ROOT,
        .mm_users               = ATOMIC_INIT(2),
        .mm_count               = ATOMIC_INIT(1),
        MMAP_LOCK_INITIALIZER(efi_mm)
index 90021d0c053095e450355d022e5b683f910d39fa..104d45ba921d4384817f879fdb6cf43887abd58c 100644 (file)
@@ -386,7 +386,6 @@ struct mm_struct {
        struct {
                struct vm_area_struct *mmap;            /* list of VMAs */
                struct maple_tree mm_mt;
-               struct rb_root mm_rb;
                u64 vmacache_seqnum;                   /* per-thread vmacache */
 #ifdef CONFIG_MMU
                unsigned long (*get_unmapped_area) (struct file *filp,
index 04b480e532b2fcd5cda854b3755453998c0e9795..ee4d445194412b6bdcd41882f502a78b520a83c4 100644 (file)
@@ -486,7 +486,6 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
                                        struct mm_struct *oldmm)
 {
        struct vm_area_struct *mpnt, *tmp, *prev, **pprev;
-       struct rb_node **rb_link, *rb_parent;
        int retval;
        unsigned long charge;
        MA_STATE(old_mas, &oldmm->mm_mt, 0, 0);
@@ -512,8 +511,6 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
        mm->exec_vm = oldmm->exec_vm;
        mm->stack_vm = oldmm->stack_vm;
 
-       rb_link = &mm->mm_rb.rb_node;
-       rb_parent = NULL;
        pprev = &mm->mmap;
        retval = ksm_fork(mm, oldmm);
        if (retval)
@@ -609,10 +606,6 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
                tmp->vm_prev = prev;
                prev = tmp;
 
-               __vma_link_rb(mm, tmp, rb_link, rb_parent);
-               rb_link = &tmp->vm_rb.rb_right;
-               rb_parent = &tmp->vm_rb;
-
                /* Link the vma into the MT */
                vma_store(mm, tmp);
 
@@ -1034,7 +1027,6 @@ static struct mm_struct *mm_init(struct mm_struct *mm, struct task_struct *p,
        struct user_namespace *user_ns)
 {
        mm->mmap = NULL;
-       mm->mm_rb = RB_ROOT;
        mt_init_flags(&mm->mm_mt, MAPLE_ALLOC_RANGE);
        mm->vmacache_seqnum = 0;
        atomic_set(&mm->mm_users, 1);
index 27229044a070211f08e930c1070d8a6cc6d26572..49d803fc4cdeaa5bf6cb34a9563b22ef9d513c6f 100644 (file)
@@ -1,6 +1,5 @@
 // SPDX-License-Identifier: GPL-2.0
 #include <linux/mm_types.h>
-#include <linux/rbtree.h>
 #include <linux/maple_tree.h>
 #include <linux/rwsem.h>
 #include <linux/spinlock.h>
@@ -28,7 +27,6 @@
  * and size this cpu_bitmask to NR_CPUS.
  */
 struct mm_struct init_mm = {
-       .mm_rb          = RB_ROOT,
        .mm_mt          = MTREE_INIT(mm_mt, MAPLE_ALLOC_RANGE),
        .pgd            = swapper_pg_dir,
        .mm_users       = ATOMIC_INIT(2),
index c2d1fac3ba5387613ff08713742ef3df74f48d00..50cbb1150e0b130e26caaccef671f86420647ad4 100644 (file)
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -38,7 +38,6 @@
 #include <linux/audit.h>
 #include <linux/khugepaged.h>
 #include <linux/uprobes.h>
-#include <linux/rbtree_augmented.h>
 #include <linux/notifier.h>
 #include <linux/memory.h>
 #include <linux/printk.h>
@@ -289,93 +288,6 @@ out:
        return retval;
 }
 
-static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
-{
-       unsigned long gap, prev_end;
-
-       /*
-        * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
-        * allow two stack_guard_gaps between them here, and when choosing
-        * an unmapped area; whereas when expanding we only require one.
-        * That's a little inconsistent, but keeps the code here simpler.
-        */
-       gap = vm_start_gap(vma);
-       if (vma->vm_prev) {
-               prev_end = vm_end_gap(vma->vm_prev);
-               if (gap > prev_end)
-                       gap -= prev_end;
-               else
-                       gap = 0;
-       }
-       return gap;
-}
-
-#ifdef CONFIG_DEBUG_VM_RB
-static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
-{
-       unsigned long max = vma_compute_gap(vma), subtree_gap;
-       if (vma->vm_rb.rb_left) {
-               subtree_gap = rb_entry(vma->vm_rb.rb_left,
-                               struct vm_area_struct, vm_rb)->rb_subtree_gap;
-               if (subtree_gap > max)
-                       max = subtree_gap;
-       }
-       if (vma->vm_rb.rb_right) {
-               subtree_gap = rb_entry(vma->vm_rb.rb_right,
-                               struct vm_area_struct, vm_rb)->rb_subtree_gap;
-               if (subtree_gap > max)
-                       max = subtree_gap;
-       }
-       return max;
-}
-
-static int browse_rb(struct mm_struct *mm)
-{
-       struct rb_root *root = &mm->mm_rb;
-       int i = 0, j, bug = 0;
-       struct rb_node *nd, *pn = NULL;
-       unsigned long prev = 0, pend = 0;
-
-       for (nd = rb_first(root); nd; nd = rb_next(nd)) {
-               struct vm_area_struct *vma;
-               vma = rb_entry(nd, struct vm_area_struct, vm_rb);
-               if (vma->vm_start < prev) {
-                       pr_emerg("vm_start %lx < prev %lx\n",
-                                 vma->vm_start, prev);
-                       bug = 1;
-               }
-               if (vma->vm_start < pend) {
-                       pr_emerg("vm_start %lx < pend %lx\n",
-                                 vma->vm_start, pend);
-                       bug = 1;
-               }
-               if (vma->vm_start > vma->vm_end) {
-                       pr_emerg("vm_start %lx > vm_end %lx\n",
-                                 vma->vm_start, vma->vm_end);
-                       bug = 1;
-               }
-               spin_lock(&mm->page_table_lock);
-               if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
-                       pr_emerg("free gap %lx, correct %lx\n",
-                              vma->rb_subtree_gap,
-                              vma_compute_subtree_gap(vma));
-                       bug = 1;
-               }
-               spin_unlock(&mm->page_table_lock);
-               i++;
-               pn = nd;
-               prev = vma->vm_start;
-               pend = vma->vm_end;
-       }
-       j = 0;
-       for (nd = pn; nd; nd = rb_prev(nd))
-               j++;
-       if (i != j) {
-               pr_emerg("backwards %d, forwards %d\n", j, i);
-               bug = 1;
-       }
-       return bug ? -1 : i;
-}
 #if defined(CONFIG_DEBUG_MAPLE_TREE)
 extern void mt_dump(const struct maple_tree *mt);
 
@@ -403,17 +315,25 @@ static void validate_mm_mt(struct mm_struct *mm)
                        dump_stack();
 #ifdef CONFIG_DEBUG_VM
                        dump_vma(vma_mt);
-                       pr_emerg("and next in rb\n");
+                       pr_emerg("and vm_next\n");
                        dump_vma(vma->vm_next);
-#endif
+#endif // CONFIG_DEBUG_VM
                        pr_emerg("mt piv: %px %lu - %lu\n", vma_mt,
                                 mas.index, mas.last);
                        pr_emerg("mt vma: %px %lu - %lu\n", vma_mt,
                                 vma_mt->vm_start, vma_mt->vm_end);
-                       pr_emerg("rb vma: %px %lu - %lu\n", vma,
+                       if (vma->vm_prev) {
+                               pr_emerg("ll prev: %px %lu - %lu\n",
+                                        vma->vm_prev, vma->vm_prev->vm_start,
+                                        vma->vm_prev->vm_end);
+                       }
+                       pr_emerg("ll vma: %px %lu - %lu\n", vma,
                                 vma->vm_start, vma->vm_end);
-                       pr_emerg("rb->next = %px %lu - %lu\n", vma->vm_next,
-                                       vma->vm_next->vm_start, vma->vm_next->vm_end);
+                       if (vma->vm_next) {
+                               pr_emerg("ll next: %px %lu - %lu\n",
+                                        vma->vm_next, vma->vm_next->vm_start,
+                                        vma->vm_next->vm_end);
+                       }
 
                        mt_dump(mas.tree);
                        if (vma_mt->vm_end != mas.last + 1) {
@@ -439,20 +359,6 @@ static void validate_mm_mt(struct mm_struct *mm)
        rcu_read_unlock();
        mt_validate(&mm->mm_mt);
 }
-#endif
-static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
-{
-       struct rb_node *nd;
-
-       for (nd = rb_first(root); nd; nd = rb_next(nd)) {
-               struct vm_area_struct *vma;
-               vma = rb_entry(nd, struct vm_area_struct, vm_rb);
-               VM_BUG_ON_VMA(vma != ignore &&
-                       vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
-                       vma);
-       }
-}
-
 static void validate_mm(struct mm_struct *mm)
 {
        int bug = 0;
@@ -460,6 +366,8 @@ static void validate_mm(struct mm_struct *mm)
        unsigned long highest_address = 0;
        struct vm_area_struct *vma = mm->mmap;
 
+       validate_mm_mt(mm);
+
        while (vma) {
                struct anon_vma *anon_vma = vma->anon_vma;
                struct anon_vma_chain *avc;
@@ -484,82 +392,12 @@ static void validate_mm(struct mm_struct *mm)
                          mm->highest_vm_end, highest_address);
                bug = 1;
        }
-       i = browse_rb(mm);
-       if (i != mm->map_count) {
-               if (i != -1)
-                       pr_emerg("map_count %d rb %d\n", mm->map_count, i);
-               bug = 1;
-       }
        VM_BUG_ON_MM(bug, mm);
 }
-#else
-#define validate_mm_rb(root, ignore) do { } while (0)
+#else // !CONFIG_DEBUG_MAPLE_TREE
 #define validate_mm_mt(root) do { } while (0)
 #define validate_mm(mm) do { } while (0)
-#endif
-
-RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
-                        struct vm_area_struct, vm_rb,
-                        unsigned long, rb_subtree_gap, vma_compute_gap)
-
-/*
- * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
- * vma->vm_prev->vm_end values changed, without modifying the vma's position
- * in the rbtree.
- */
-static void vma_gap_update(struct vm_area_struct *vma)
-{
-       /*
-        * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
-        * a callback function that does exactly what we want.
-        */
-       vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
-}
-
-static inline void vma_rb_insert(struct vm_area_struct *vma,
-                                struct rb_root *root)
-{
-       /* All rb_subtree_gap values must be consistent prior to insertion */
-       validate_mm_rb(root, NULL);
-
-       rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
-}
-
-static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
-{
-       /*
-        * Note rb_erase_augmented is a fairly large inline function,
-        * so make sure we instantiate it only once with our desired
-        * augmented rbtree callbacks.
-        */
-       rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
-}
-
-static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
-                                               struct rb_root *root,
-                                               struct vm_area_struct *ignore)
-{
-       /*
-        * All rb_subtree_gap values must be consistent prior to erase,
-        * with the possible exception of the "next" vma being erased if
-        * next->vm_start was reduced.
-        */
-       validate_mm_rb(root, ignore);
-
-       __vma_rb_erase(vma, root);
-}
-
-static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
-                                        struct rb_root *root)
-{
-       /*
-        * All rb_subtree_gap values must be consistent prior to erase,
-        * with the possible exception of the vma being erased.
-        */
-       validate_mm_rb(root, vma);
-
-       __vma_rb_erase(vma, root);
-}
+#endif // CONFIG_DEBUG_MAPLE_TREE
 
 /*
  * vma has some anon_vma assigned, and is already inserted on that
@@ -593,38 +431,29 @@ anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
                anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
 }
 
-static int find_vma_links(struct mm_struct *mm, unsigned long addr,
-               unsigned long end, struct vm_area_struct **pprev,
-               struct rb_node ***rb_link, struct rb_node **rb_parent)
+/* Private
+ * range_has_overlap() - Check the @start - @end range for overlapping VMAs and
+ * sets up a pointer to the previous VMA
+ *
+ * @mm - the mm struct
+ * @start - the start address of the range
+ * @end - the end address of the range
+ * @pprev - the pointer to the pointer of the previous VMA
+ *
+ * Returns: True if there is an overlapping VMA, false otherwise
+ */
+static bool range_has_overlap(struct mm_struct *mm, unsigned long start,
+                             unsigned long end, struct vm_area_struct **pprev)
 {
-       struct rb_node **__rb_link, *__rb_parent, *rb_prev;
-
-       __rb_link = &mm->mm_rb.rb_node;
-       rb_prev = __rb_parent = NULL;
-
-       while (*__rb_link) {
-               struct vm_area_struct *vma_tmp;
+       struct vm_area_struct *existing;
 
-               __rb_parent = *__rb_link;
-               vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
-
-               if (vma_tmp->vm_end > addr) {
-                       /* Fail if an existing vma overlaps the area */
-                       if (vma_tmp->vm_start < end)
-                               return -ENOMEM;
-                       __rb_link = &__rb_parent->rb_left;
-               } else {
-                       rb_prev = __rb_parent;
-                       __rb_link = &__rb_parent->rb_right;
-               }
-       }
+       MA_STATE(mas, &mm->mm_mt, start, start);
+       existing = mas_find(&mas, end - 1);
+       if (mas.node == MAS_NONE)
+               mas_reset(&mas);
 
-       *pprev = NULL;
-       if (rb_prev)
-               *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
-       *rb_link = __rb_link;
-       *rb_parent = __rb_parent;
-       return 0;
+       *pprev = mas_prev(&mas, 0);
+       return existing ? true : false;
 }
 
 /* Private
@@ -655,83 +484,43 @@ static inline struct vm_area_struct *next_vma_or_first(struct mm_struct *mm,
  * @start: The start address.
  * @len: The length of the range.
  * @pprev: pointer to the pointer that will be set to previous vm_area_struct
- * @rb_link - the rb_node
- * @rb_parent - the parent rb_node
  *
  * 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 rb_node ***rb_link,
-                struct rb_node **rb_parent, struct list_head *uf)
+                struct vm_area_struct **pprev, struct list_head *uf)
 {
-
-       while (find_vma_links(mm, start, start + len, pprev, rb_link, rb_parent))
+       // 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)
 {
        unsigned long nr_pages = 0;
-       unsigned long nr_mt_pages = 0;
        struct vm_area_struct *vma;
+       unsigned long vm_start, vm_end;
 
        /* Find first overlaping mapping */
        vma = find_vma_intersection(mm, addr, end);
        if (!vma)
                return 0;
 
-       nr_pages = (min(end, vma->vm_end) -
-               max(addr, vma->vm_start)) >> PAGE_SHIFT;
-
-       /* Iterate over the rest of the overlaps */
-       for (vma = vma->vm_next; vma; vma = vma->vm_next) {
-               unsigned long overlap_len;
-
-               if (vma->vm_start > end)
-                       break;
-
-               overlap_len = min(end, vma->vm_end) - vma->vm_start;
-               nr_pages += overlap_len >> PAGE_SHIFT;
-       }
+       vm_start = vma->vm_start;
+       vm_end = vma->vm_end;
 
-       mt_for_each(&mm->mm_mt, vma, addr, end) {
-               nr_mt_pages +=
-                       (min(end, vma->vm_end) - vma->vm_start) >> PAGE_SHIFT;
-       }
+       nr_pages = (min(end, vm_end) - max(addr, vm_start)) >> PAGE_SHIFT;
 
-       VM_BUG_ON_MM(nr_pages != nr_mt_pages, mm);
+       /* Iterate over the rest of the overlaps */
+       mt_for_each(&mm->mm_mt, vma, addr, end)
+               nr_pages += (min(end, vm_end) - vm_start) >> PAGE_SHIFT;
 
        return nr_pages;
 }
 
-void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
-               struct rb_node **rb_link, struct rb_node *rb_parent)
-{
-       /* Update tracking information for the gap following the new vma. */
-       if (vma->vm_next)
-               vma_gap_update(vma->vm_next);
-       else
-               mm->highest_vm_end = vm_end_gap(vma);
-
-       /*
-        * vma->vm_prev wasn't known when we followed the rbtree to find the
-        * correct insertion point for that vma. As a result, we could not
-        * update the vma vm_rb parents rb_subtree_gap values on the way down.
-        * So, we first insert the vma with a zero rb_subtree_gap value
-        * (to be consistent with what we did on the way down), and then
-        * immediately update the gap to the correct value. Finally we
-        * rebalance the rbtree after all augmented values have been set.
-        */
-       rb_link_node(&vma->vm_rb, rb_parent, rb_link);
-       vma->rb_subtree_gap = 0;
-       vma_gap_update(vma);
-       vma_rb_insert(vma, &mm->mm_rb);
-}
-
 static void __vma_link_file(struct vm_area_struct *vma)
 {
        struct file *file;
@@ -799,17 +588,14 @@ void vma_store(struct mm_struct *mm, struct vm_area_struct *vma)
 
 static void
 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
-       struct vm_area_struct *prev, struct rb_node **rb_link,
-       struct rb_node *rb_parent)
+       struct vm_area_struct *prev)
 {
        vma_mt_store(mm, vma);
        __vma_link_list(mm, vma, prev);
-       __vma_link_rb(mm, vma, rb_link, rb_parent);
 }
 
 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
-                       struct vm_area_struct *prev, struct rb_node **rb_link,
-                       struct rb_node *rb_parent)
+                       struct vm_area_struct *prev)
 {
        struct address_space *mapping = NULL;
 
@@ -818,7 +604,7 @@ static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
                i_mmap_lock_write(mapping);
        }
 
-       __vma_link(mm, vma, prev, rb_link, rb_parent);
+       __vma_link(mm, vma, prev);
        __vma_link_file(vma);
 
        if (mapping)
@@ -830,30 +616,18 @@ static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 
 /*
  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
- * mm's list and rbtree.  It has already been inserted into the interval tree.
+ * mm's list and the mm tree.  It has already been inserted into the interval tree.
  */
 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
 {
        struct vm_area_struct *prev;
-       struct rb_node **rb_link, *rb_parent;
 
-       if (find_vma_links(mm, vma->vm_start, vma->vm_end,
-                          &prev, &rb_link, &rb_parent))
+       if (range_has_overlap(mm, vma->vm_start, vma->vm_end, &prev))
                BUG();
-       __vma_link(mm, vma, prev, rb_link, rb_parent);
+       __vma_link(mm, vma, prev);
        mm->map_count++;
 }
 
-static __always_inline void __vma_unlink_common(struct mm_struct *mm,
-                                               struct vm_area_struct *vma,
-                                               struct vm_area_struct *ignore)
-{
-       vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
-       __vma_unlink_list(mm, vma);
-       /* Kill the cache */
-       vmacache_invalidate(mm);
-}
-
 /*
  * 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.
@@ -875,9 +649,6 @@ int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
        long adjust_next = 0;
        int remove_next = 0;
 
-       validate_mm(mm);
-       validate_mm_mt(mm);
-
        if (next && !insert) {
                struct vm_area_struct *exporter = NULL, *importer = NULL;
 
@@ -1038,25 +809,9 @@ again:
        }
 
        if (remove_next) {
-               /*
-                * vma_merge has merged next into vma, and needs
-                * us to remove next before dropping the locks.
-                * Since we have expanded over this vma, the maple tree will
-                * have overwritten by storing the value
-                */
-               if (remove_next != 3)
-                       __vma_unlink_common(mm, next, next);
-               else
-                       /*
-                        * vma is not before next if they've been
-                        * swapped.
-                        *
-                        * pre-swap() next->vm_start was reduced so
-                        * tell validate_mm_rb to ignore pre-swap()
-                        * "next" (which is stored in post-swap()
-                        * "vma").
-                        */
-                       __vma_unlink_common(mm, next, vma);
+               __vma_unlink_list(mm, next);
+               /* Kill the cache */
+               vmacache_invalidate(mm);
                if (file)
                        __remove_shared_vm_struct(next, file, mapping);
        } else if (insert) {
@@ -1066,15 +821,8 @@ again:
                 * (it may either follow vma or precede it).
                 */
                __insert_vm_struct(mm, insert);
-       } else {
-               if (start_changed)
-                       vma_gap_update(vma);
-               if (end_changed) {
-                       if (!next)
-                               mm->highest_vm_end = vm_end_gap(vma);
-                       else if (!adjust_next)
-                               vma_gap_update(next);
-               }
+       } else if (end_changed && !next) {
+               mm->highest_vm_end = vm_end_gap(vma);
        }
 
        if (anon_vma) {
@@ -1133,10 +881,7 @@ again:
                        remove_next = 1;
                        end = next->vm_end;
                        goto again;
-               }
-               else if (next)
-                       vma_gap_update(next);
-               else {
+               } else if (!next) {
                        /*
                         * If remove_next == 2 we obviously can't
                         * reach this path.
@@ -1163,8 +908,6 @@ again:
                uprobe_mmap(insert);
 
        validate_mm(mm);
-       validate_mm_mt(mm);
-
        return 0;
 }
 
@@ -1311,7 +1054,6 @@ struct vm_area_struct *vma_merge(struct mm_struct *mm,
        struct vm_area_struct *area, *next;
        int err;
 
-       validate_mm_mt(mm);
        /*
         * We later require that vma->vm_flags == vm_flags,
         * so this tests vma->vm_flags & VM_SPECIAL, too.
@@ -1387,7 +1129,6 @@ struct vm_area_struct *vma_merge(struct mm_struct *mm,
                khugepaged_enter_vma_merge(area, vm_flags);
                return area;
        }
-       validate_mm_mt(mm);
 
        return NULL;
 }
@@ -1558,6 +1299,7 @@ unsigned long do_mmap(struct file *file, unsigned long addr,
        struct mm_struct *mm = current->mm;
        int pkey = 0;
 
+       validate_mm(mm);
        *populate = 0;
 
        if (!len)
@@ -1877,10 +1619,8 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
        struct mm_struct *mm = current->mm;
        struct vm_area_struct *vma, *prev;
        int error;
-       struct rb_node **rb_link, *rb_parent;
        unsigned long charged = 0;
 
-       validate_mm_mt(mm);
        /* Check against address space limit. */
        if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
                unsigned long nr_pages;
@@ -1896,8 +1636,8 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
                        return -ENOMEM;
        }
 
-       /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
-       if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
+       /* Clear old maps, set up prev and uf */
+       if (munmap_vma_range(mm, addr, len, &prev, uf))
                return -ENOMEM;
        /*
         * Private writable mapping: check memory availability
@@ -1960,8 +1700,8 @@ 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()
+                * Bug: If addr is changed, prev and the maple tree data  should
+                * be updated for vma_link()
                 */
                WARN_ON_ONCE(addr != vma->vm_start);
 
@@ -1975,7 +1715,7 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
                vma_set_anonymous(vma);
        }
 
-       vma_link(mm, vma, prev, rb_link, rb_parent);
+       vma_link(mm, vma, prev);
        /* Once vma denies write, undo our temporary denial count */
        if (file) {
                if (vm_flags & VM_SHARED)
@@ -2011,7 +1751,6 @@ out:
 
        vma_set_page_prot(vma);
 
-       validate_mm_mt(mm);
        return addr;
 
 unmap_and_free_vma:
@@ -2031,7 +1770,6 @@ free_vma:
 unacct_error:
        if (charged)
                vm_unacct_memory(charged);
-       validate_mm_mt(mm);
        return error;
 }
 
@@ -2049,7 +1787,6 @@ static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
        unsigned long length, gap;
 
        MA_STATE(mas, &current->mm->mm_mt, 0, 0);
-       validate_mm(current->mm);
 
        /* Adjust search length to account for worst case alignment overhead */
        length = info->length + info->align_mask;
@@ -2082,7 +1819,6 @@ static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
        unsigned long length, gap;
 
        MA_STATE(mas, &current->mm->mm_mt, 0, 0);
-       validate_mm_mt(current->mm);
        /* Adjust search length to account for worst case alignment overhead */
        length = info->length + info->align_mask;
        if (length < info->length)
@@ -2323,7 +2059,7 @@ find_vma_prev(struct mm_struct *mm, unsigned long addr,
 
        rcu_read_lock();
        vma = mas_find(&mas, ULONG_MAX);
-       if (!vma)
+       if (mas.node == MAS_NONE)
                mas_reset(&mas);
 
        *pprev = mas_prev(&mas, 0);
@@ -2389,7 +2125,6 @@ int expand_upwards(struct vm_area_struct *vma, unsigned long address)
        unsigned long gap_addr;
        int error = 0;
 
-       validate_mm_mt(mm);
        if (!(vma->vm_flags & VM_GROWSUP))
                return -EFAULT;
 
@@ -2436,15 +2171,13 @@ int expand_upwards(struct vm_area_struct *vma, unsigned long address)
                        error = acct_stack_growth(vma, size, grow);
                        if (!error) {
                                /*
-                                * vma_gap_update() doesn't support concurrent
-                                * updates, but we only hold a shared mmap_lock
-                                * lock here, so we need to protect against
-                                * concurrent vma expansions.
-                                * anon_vma_lock_write() doesn't help here, as
-                                * we don't guarantee that all growable vmas
-                                * in a mm share the same root anon vma.
-                                * So, we reuse mm->page_table_lock to guard
-                                * against concurrent vma expansions.
+                                * We only hold a shared mmap_lock lock here, so
+                                * we need to protect against concurrent vma
+                                * expansions.  anon_vma_lock_write() doesn't
+                                * help here, as we don't guarantee that all
+                                * growable vmas in a mm share the same root
+                                * anon vma.  So, we reuse mm->page_table_lock
+                                * to guard against concurrent vma expansions.
                                 */
                                spin_lock(&mm->page_table_lock);
                                if (vma->vm_flags & VM_LOCKED)
@@ -2452,10 +2185,9 @@ int expand_upwards(struct vm_area_struct *vma, unsigned long address)
                                vm_stat_account(mm, vma->vm_flags, grow);
                                anon_vma_interval_tree_pre_update_vma(vma);
                                vma->vm_end = address;
+                               vma_mt_store(mm, vma);
                                anon_vma_interval_tree_post_update_vma(vma);
-                               if (vma->vm_next)
-                                       vma_gap_update(vma->vm_next);
-                               else
+                               if (!vma->vm_next)
                                        mm->highest_vm_end = vm_end_gap(vma);
                                spin_unlock(&mm->page_table_lock);
 
@@ -2466,7 +2198,6 @@ int expand_upwards(struct vm_area_struct *vma, unsigned long address)
        anon_vma_unlock_write(vma->anon_vma);
        khugepaged_enter_vma_merge(vma, vma->vm_flags);
        validate_mm(mm);
-       validate_mm_mt(mm);
        return error;
 }
 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
@@ -2481,7 +2212,6 @@ int expand_downwards(struct vm_area_struct *vma,
        struct vm_area_struct *prev;
        int error = 0;
 
-       validate_mm(mm);
        address &= PAGE_MASK;
        if (address < mmap_min_addr)
                return -EPERM;
@@ -2518,15 +2248,13 @@ int expand_downwards(struct vm_area_struct *vma,
                        error = acct_stack_growth(vma, size, grow);
                        if (!error) {
                                /*
-                                * vma_gap_update() doesn't support concurrent
-                                * updates, but we only hold a shared mmap_lock
-                                * lock here, so we need to protect against
-                                * concurrent vma expansions.
-                                * anon_vma_lock_write() doesn't help here, as
-                                * we don't guarantee that all growable vmas
-                                * in a mm share the same root anon vma.
-                                * So, we reuse mm->page_table_lock to guard
-                                * against concurrent vma expansions.
+                                * We only hold a shared mmap_lock lock here, so
+                                * we need to protect against concurrent vma
+                                * expansions.  anon_vma_lock_write() doesn't
+                                * help here, as we don't guarantee that all
+                                * growable vmas in a mm share the same root
+                                * anon vma.  So, we reuse mm->page_table_lock
+                                * to guard against concurrent vma expansions.
                                 */
                                spin_lock(&mm->page_table_lock);
                                if (vma->vm_flags & VM_LOCKED)
@@ -2538,7 +2266,6 @@ int expand_downwards(struct vm_area_struct *vma,
                                /* Overwrite old entry in mtree. */
                                vma_mt_store(mm, vma);
                                anon_vma_interval_tree_post_update_vma(vma);
-                               vma_gap_update(vma);
                                spin_unlock(&mm->page_table_lock);
 
                                perf_event_mmap(vma);
@@ -2682,17 +2409,15 @@ detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
        insertion_point = (prev ? &prev->vm_next : &mm->mmap);
        vma->vm_prev = NULL;
        do {
-               vma_rb_erase(vma, &mm->mm_rb);
                vma_mt_erase(mm, vma);
                mm->map_count--;
                tail_vma = vma;
                vma = vma->vm_next;
        } while (vma && vma->vm_start < end);
        *insertion_point = vma;
-       if (vma) {
+       if (vma)
                vma->vm_prev = prev;
-               vma_gap_update(vma);
-       } else
+       else
                mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
        tail_vma->vm_next = NULL;
 
@@ -2798,11 +2523,7 @@ int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
        if (len == 0)
                return -EINVAL;
 
-       /*
-        * arch_unmap() might do unmaps itself.  It must be called
-        * and finish any rbtree manipulation before this code
-        * runs and also starts to manipulate the rbtree.
-        */
+        /* arch_unmap() might do unmaps itself.  */
        arch_unmap(mm, start, end);
 
        /* Find the first overlapping VMA */
@@ -2810,7 +2531,7 @@ int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
        if (!vma)
                return 0;
        prev = vma->vm_prev;
-       /* we have  start < vma->vm_end  */
+       /* we have start < vma->vm_end  */
 
        /* if it doesn't overlap, we have nothing.. */
        if (vma->vm_start >= end)
@@ -2879,7 +2600,7 @@ int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
                }
        }
 
-       /* Detach vmas from rbtree */
+       /* Detach vmas from the MM linked list and remove from the mm tree*/
        detach_vmas_to_be_unmapped(mm, vma, prev, end);
 
        if (downgrade)
@@ -3042,11 +2763,11 @@ out:
  *  anonymous maps.  eventually we may be able to do some
  *  brk-specific accounting here.
  */
-static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
+static int do_brk_flags(unsigned long addr, unsigned long len,
+                       unsigned long flags, struct list_head *uf)
 {
        struct mm_struct *mm = current->mm;
        struct vm_area_struct *vma, *prev;
-       struct rb_node **rb_link, *rb_parent;
        pgoff_t pgoff = addr >> PAGE_SHIFT;
        int error;
        unsigned long mapped_addr;
@@ -3065,8 +2786,8 @@ static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long fla
        if (error)
                return error;
 
-       /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
-       if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
+       /* Clear old maps, set up prev and uf */
+       if (munmap_vma_range(mm, addr, len, &prev, uf))
                return -ENOMEM;
 
        /* Check against address space limits *after* clearing old maps... */
@@ -3100,7 +2821,7 @@ static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long fla
        vma->vm_pgoff = pgoff;
        vma->vm_flags = flags;
        vma->vm_page_prot = vm_get_page_prot(flags);
-       vma_link(mm, vma, prev, rb_link, rb_parent);
+       vma_link(mm, vma, prev);
 out:
        perf_event_mmap(vma);
        mm->total_vm += len >> PAGE_SHIFT;
@@ -3225,26 +2946,10 @@ void exit_mmap(struct mm_struct *mm)
 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
 {
        struct vm_area_struct *prev;
-       struct rb_node **rb_link, *rb_parent;
-       unsigned long start = vma->vm_start;
-       struct vm_area_struct *overlap = NULL;
 
-       if (find_vma_links(mm, vma->vm_start, vma->vm_end,
-                          &prev, &rb_link, &rb_parent))
+       if (range_has_overlap(mm, vma->vm_start, vma->vm_end, &prev))
                return -ENOMEM;
 
-       overlap = mt_find(&mm->mm_mt, &start, vma->vm_end - 1);
-       if (overlap) {
-
-               pr_err("Found vma ending at %lu\n", start - 1);
-               pr_err("vma : %lu => %lu-%lu\n", (unsigned long)overlap,
-                               overlap->vm_start, overlap->vm_end - 1);
-#if defined(CONFIG_DEBUG_MAPLE_TREE)
-               mt_dump(&mm->mm_mt);
-#endif
-               BUG();
-       }
-
        if ((vma->vm_flags & VM_ACCOUNT) &&
             security_vm_enough_memory_mm(mm, vma_pages(vma)))
                return -ENOMEM;
@@ -3266,7 +2971,7 @@ int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
                vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
        }
 
-       vma_link(mm, vma, prev, rb_link, rb_parent);
+       vma_link(mm, vma, prev);
        return 0;
 }
 
@@ -3282,9 +2987,7 @@ struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
        unsigned long vma_start = vma->vm_start;
        struct mm_struct *mm = vma->vm_mm;
        struct vm_area_struct *new_vma, *prev;
-       struct rb_node **rb_link, *rb_parent;
        bool faulted_in_anon_vma = true;
-       unsigned long index = addr;
 
        validate_mm_mt(mm);
        /*
@@ -3296,10 +2999,9 @@ struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
                faulted_in_anon_vma = false;
        }
 
-       if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
+       if (range_has_overlap(mm, addr, addr + len, &prev))
                return NULL;    /* should never get here */
-       if (mt_find(&mm->mm_mt, &index, addr+len - 1))
-               BUG();
+
        new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
                            vma->vm_userfaultfd_ctx);
@@ -3340,7 +3042,7 @@ struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
                        get_file(new_vma->vm_file);
                if (new_vma->vm_ops && new_vma->vm_ops->open)
                        new_vma->vm_ops->open(new_vma);
-               vma_link(mm, new_vma, prev, rb_link, rb_parent);
+               vma_link(mm, new_vma, prev);
                *need_rmap_locks = false;
        }
        validate_mm_mt(mm);
index c63c8e47be572d7e8349eb9f04d2dcf7c3601e5c..c9fae684912fe88f57bd2a96c71e6d85b29cc222 100644 (file)
--- a/mm/util.c
+++ b/mm/util.c
@@ -286,6 +286,8 @@ void __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
        vma->vm_next = next;
        if (next)
                next->vm_prev = vma;
+       else
+               mm->highest_vm_end = vm_end_gap(vma);
 }
 
 void __vma_unlink_list(struct mm_struct *mm, struct vm_area_struct *vma)
@@ -300,6 +302,12 @@ void __vma_unlink_list(struct mm_struct *mm, struct vm_area_struct *vma)
                mm->mmap = next;
        if (next)
                next->vm_prev = prev;
+       else {
+               if (prev)
+                       mm->highest_vm_end = vm_end_gap(prev);
+               else
+                       mm->highest_vm_end = 0;
+       }
 }
 
 /* Check if the vma is being used as a stack by this task */