From d72da4a4d973d8a0a0d3c97e7cdebf287fbe3a99 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Wed, 27 May 2015 11:09:36 +0930 Subject: rbtree: Make lockless searches non-fatal Change the insert and erase code such that lockless searches are non-fatal. In and of itself an rbtree cannot be correctly searched while in-modification, we can however provide weaker guarantees that will allow the rbtree to be used in conjunction with other techniques, such as latches; see 9b0fd802e8c0 ("seqcount: Add raw_write_seqcount_latch()"). For this to work we need the following guarantees from the rbtree code: 1) a lockless reader must not see partial stores, this would allow it to observe nodes that are invalid memory. 2) there must not be (temporary) loops in the tree structure in the modifier's program order, this would cause a lookup which interrupts the modifier to get stuck indefinitely. For 1) we must use WRITE_ONCE() for all updates to the tree structure; in particular this patch only does rb_{left,right} as those are the only element required for simple searches. It generates slightly worse code, probably because volatile. But in pointer chasing heavy code a few instructions more should not matter. For 2) I have carefully audited the code and drawn every intermediate link state and not found a loop. Cc: Mathieu Desnoyers Cc: "Paul E. McKenney" Cc: Oleg Nesterov Cc: Andrea Arcangeli Cc: David Woodhouse Cc: Rik van Riel Reviewed-by: Michel Lespinasse Signed-off-by: Peter Zijlstra (Intel) Signed-off-by: Rusty Russell --- lib/rbtree.c | 76 ++++++++++++++++++++++++++++++++++++++++++------------------ 1 file changed, 54 insertions(+), 22 deletions(-) (limited to 'lib/rbtree.c') diff --git a/lib/rbtree.c b/lib/rbtree.c index c16c81a3d430..1356454e36de 100644 --- a/lib/rbtree.c +++ b/lib/rbtree.c @@ -44,6 +44,30 @@ * parentheses and have some accompanying text comment. */ +/* + * Notes on lockless lookups: + * + * All stores to the tree structure (rb_left and rb_right) must be done using + * WRITE_ONCE(). And we must not inadvertently cause (temporary) loops in the + * tree structure as seen in program order. + * + * These two requirements will allow lockless iteration of the tree -- not + * correct iteration mind you, tree rotations are not atomic so a lookup might + * miss entire subtrees. + * + * But they do guarantee that any such traversal will only see valid elements + * and that it will indeed complete -- does not get stuck in a loop. + * + * It also guarantees that if the lookup returns an element it is the 'correct' + * one. But not returning an element does _NOT_ mean it's not present. + * + * NOTE: + * + * Stores to __rb_parent_color are not important for simple lookups so those + * are left undone as of now. Nor did I check for loops involving parent + * pointers. + */ + static inline void rb_set_black(struct rb_node *rb) { rb->__rb_parent_color |= RB_BLACK; @@ -129,8 +153,9 @@ __rb_insert(struct rb_node *node, struct rb_root *root, * This still leaves us in violation of 4), the * continuation into Case 3 will fix that. */ - parent->rb_right = tmp = node->rb_left; - node->rb_left = parent; + tmp = node->rb_left; + WRITE_ONCE(parent->rb_right, tmp); + WRITE_ONCE(node->rb_left, parent); if (tmp) rb_set_parent_color(tmp, parent, RB_BLACK); @@ -149,8 +174,8 @@ __rb_insert(struct rb_node *node, struct rb_root *root, * / \ * n U */ - gparent->rb_left = tmp; /* == parent->rb_right */ - parent->rb_right = gparent; + WRITE_ONCE(gparent->rb_left, tmp); /* == parent->rb_right */ + WRITE_ONCE(parent->rb_right, gparent); if (tmp) rb_set_parent_color(tmp, gparent, RB_BLACK); __rb_rotate_set_parents(gparent, parent, root, RB_RED); @@ -171,8 +196,9 @@ __rb_insert(struct rb_node *node, struct rb_root *root, tmp = parent->rb_left; if (node == tmp) { /* Case 2 - right rotate at parent */ - parent->rb_left = tmp = node->rb_right; - node->rb_right = parent; + tmp = node->rb_right; + WRITE_ONCE(parent->rb_left, tmp); + WRITE_ONCE(node->rb_right, parent); if (tmp) rb_set_parent_color(tmp, parent, RB_BLACK); @@ -183,8 +209,8 @@ __rb_insert(struct rb_node *node, struct rb_root *root, } /* Case 3 - left rotate at gparent */ - gparent->rb_right = tmp; /* == parent->rb_left */ - parent->rb_left = gparent; + WRITE_ONCE(gparent->rb_right, tmp); /* == parent->rb_left */ + WRITE_ONCE(parent->rb_left, gparent); if (tmp) rb_set_parent_color(tmp, gparent, RB_BLACK); __rb_rotate_set_parents(gparent, parent, root, RB_RED); @@ -224,8 +250,9 @@ ____rb_erase_color(struct rb_node *parent, struct rb_root *root, * / \ / \ * Sl Sr N Sl */ - parent->rb_right = tmp1 = sibling->rb_left; - sibling->rb_left = parent; + tmp1 = sibling->rb_left; + WRITE_ONCE(parent->rb_right, tmp1); + WRITE_ONCE(sibling->rb_left, parent); rb_set_parent_color(tmp1, parent, RB_BLACK); __rb_rotate_set_parents(parent, sibling, root, RB_RED); @@ -275,9 +302,10 @@ ____rb_erase_color(struct rb_node *parent, struct rb_root *root, * \ * Sr */ - sibling->rb_left = tmp1 = tmp2->rb_right; - tmp2->rb_right = sibling; - parent->rb_right = tmp2; + tmp1 = tmp2->rb_right; + WRITE_ONCE(sibling->rb_left, tmp1); + WRITE_ONCE(tmp2->rb_right, sibling); + WRITE_ONCE(parent->rb_right, tmp2); if (tmp1) rb_set_parent_color(tmp1, sibling, RB_BLACK); @@ -297,8 +325,9 @@ ____rb_erase_color(struct rb_node *parent, struct rb_root *root, * / \ / \ * (sl) sr N (sl) */ - parent->rb_right = tmp2 = sibling->rb_left; - sibling->rb_left = parent; + tmp2 = sibling->rb_left; + WRITE_ONCE(parent->rb_right, tmp2); + WRITE_ONCE(sibling->rb_left, parent); rb_set_parent_color(tmp1, sibling, RB_BLACK); if (tmp2) rb_set_parent(tmp2, parent); @@ -310,8 +339,9 @@ ____rb_erase_color(struct rb_node *parent, struct rb_root *root, sibling = parent->rb_left; if (rb_is_red(sibling)) { /* Case 1 - right rotate at parent */ - parent->rb_left = tmp1 = sibling->rb_right; - sibling->rb_right = parent; + tmp1 = sibling->rb_right; + WRITE_ONCE(parent->rb_left, tmp1); + WRITE_ONCE(sibling->rb_right, parent); rb_set_parent_color(tmp1, parent, RB_BLACK); __rb_rotate_set_parents(parent, sibling, root, RB_RED); @@ -336,9 +366,10 @@ ____rb_erase_color(struct rb_node *parent, struct rb_root *root, break; } /* Case 3 - right rotate at sibling */ - sibling->rb_right = tmp1 = tmp2->rb_left; - tmp2->rb_left = sibling; - parent->rb_left = tmp2; + tmp1 = tmp2->rb_left; + WRITE_ONCE(sibling->rb_right, tmp1); + WRITE_ONCE(tmp2->rb_left, sibling); + WRITE_ONCE(parent->rb_left, tmp2); if (tmp1) rb_set_parent_color(tmp1, sibling, RB_BLACK); @@ -347,8 +378,9 @@ ____rb_erase_color(struct rb_node *parent, struct rb_root *root, sibling = tmp2; } /* Case 4 - left rotate at parent + color flips */ - parent->rb_left = tmp2 = sibling->rb_right; - sibling->rb_right = parent; + tmp2 = sibling->rb_right; + WRITE_ONCE(parent->rb_left, tmp2); + WRITE_ONCE(sibling->rb_right, parent); rb_set_parent_color(tmp1, sibling, RB_BLACK); if (tmp2) rb_set_parent(tmp2, parent); -- cgit v1.2.3-59-g8ed1b