rbtree_latch: quit searching when reaching to maximum depth
PATCH 1/2 rbtree_latch: quit searching when reaching to maximum depth
lib/rbtree.c has ensured that there is not possible to inadvertently cause (temporary) loops in the tree structure as seen in program order of the modifier. But loop is still possible to be seen in searcher due to CPU's reordering.
for example:
modifier searcher
left rotate at parentparent->rb_right is node
search to parent parent->rb_right is node +->see node->rb_left changed WRITE_ONCE(parent->rb_right, tmp);-+ | node->rb_left is parennt no smp_wmb(), some arch can | | reorder these two writes | | loop long between WRITE_ONCE(node->rb_left, parent);-+-+ parent and node | +--->finally see parent->rb_right
The long loop won't stop until the modifer's CPU flushes its writes. Too avoid it, we should limit the searching depth. There are no more than (1<<BITS_PER_LONG)-1 nodes in the tree. And the max_depth of a tree is no more than 2*lg(node_count+1), which is no mare than 2*BITS_PER_LONG.
So the serarch should stop when diving down up to
2*BITS_PER_LONG depth.
--- include/linux/rbtree_latch.h | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
include/linux/rbtree_latch.h +3 -1
@@ -107,10 +107,11 @@ __lt_find(void *key, struct latch_tree_root *ltr, int idx, int (*comp)(void *key, struct latch_tree_node *node)) {struct rb_node *node =rcu_dereference_raw (ltr->tree[idx].rb_node); +int depth =2 *BITS_PER_LONG ;struct latch_tree_node *ltn;int c; -while (node) { +while (node && depth >0 ) { ltn =__lt_from_rb (node, idx); c =comp (key, ltn);
@@ -120,6 +121,7 @@ __lt_find(void *key, struct latch_tree_root *ltr, int idx, node =rcu_dereference_raw (node->rb_right); else return ltn; + depth--; }return NULL ;
Signature · 1 line
2.20.1
PATCH 2/2 rbtree_latch: don't need to check seq when it found a node
latch_tree_find() should be protected by caller via RCU or so. When it find a node in an attempt, the node must be a valid one.
--- include/linux/rbtree_latch.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
include/linux/rbtree_latch.h +1 -1
@@ -208,7 +208,7 @@ latch_tree_find(void *key, struct latch_tree_root *root,do { seq =raw_read_seqcount_latch (&root->seq); node =__lt_find (key, root, seq &1 , ops->comp); - }while (read_seqcount_retry (&root->seq, seq)); + }while (!node &&read_seqcount_retry (&root->seq, seq));return node; }
Signature · 1 line
2.20.1
Re: [PATCH 2/2] rbtree_latch: don't need to check seq when it found a node
Quoted 21 lines · Lai Jiangshan
latch_tree_find()should be protected by caller via RCU or so. When it find a node in an attempt, the node must be a valid one.Cc Peter Zijlstra <[email protected]>Cc Paul E. McKenney <[email protected]>Cc Oleg Nesterov <[email protected]>Cc Michel Lespinasse <[email protected]>Cc Andrea Arcangeli <[email protected]>Cc David Woodhouse <[email protected]>Cc Rik van Riel <[email protected]>Cc Mathieu Desnoyers <[email protected]>Signed-off-by Lai Jiangshan <[email protected]>--- include/linux/rbtree_latch.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)include/linux/rbtree_latch.h +0
-0
include/linux/rbtree_latch.h latch_tree_find(void *key, struct latch_tree_root *root, · line 211
Quoted 5 lines
include/linux/rbtree_latch.h +1
-1@@ -208,7 +208,7 @@ latch_tree_find(void *key, struct latch_tree_root *root,do { seq =raw_read_seqcount_latch (&root->seq); node =__lt_find (key, root, seq &1 , ops->comp); - }while (read_seqcount_retry (&root->seq, seq)); + }while (!node &&read_seqcount_retry (&root->seq, seq));
So in the case where we search for key=N and race with { erase(N); insert(N) }, we can now return the old N, as opposed to the new N.
But given that is entirely subject to timing anyway, that is irrelevant. We change the boundary case in the timing.
Quoted 21 lines · Lai Jiangshan
lib/rbtree.chas ensured that there is not possible to inadvertently cause (temporary) loops in the tree structure as seen in program order of the modifier. But loop is still possible to be seen in searcher due to CPU's reordering.for example:
modifier searcherleft rotate at parent
parent->rb_rightis nodesearch to parent parent->rb_right is node +->see node->rb_left changed WRITE_ONCE(parent->rb_right, tmp);-+ | node->rb_left is parennt no smp_wmb(), some arch can | | reorder these two writes | | loop long between WRITE_ONCE(node->rb_left, parent);-+-+ parent and node | +--->finally see parent->rb_right
Quoted 2 lines
The long loop won't stop until the modifer's CPU flushes its writes. Too avoid it, we should limit the searching depth.
Cute, have you actually observed this? Did you have performance issues?
Quoted 6 lines
There are no more than (1<<BITS_PER_LONG)-1 nodes in the tree. And the
max_depthof a tree is no more than 2*lg(node_count+1), which is no mare than 2*BITS_PER_LONG.So the serarch should stop when diving down up to
2*BITS_PER_LONG depth.
Arguably you can have a larger key space, but I think due to memory constraints this limit still isn't wrong. But I do feel you need a comment with that.
Quoted 23 lines · Peter Zijlstra
Quoted 21 lines · Lai Jiangshan
lib/rbtree.chas ensured that there is not possible to inadvertently cause (temporary) loops in the tree structure as seen in program order of the modifier. But loop is still possible to be seen in searcher due to CPU's reordering.for example:
modifier searcherleft rotate at parent
parent->rb_rightis nodesearch to parent parent->rb_right is node +->see node->rb_left changed WRITE_ONCE(parent->rb_right, tmp);-+ | node->rb_left is parennt no smp_wmb(), some arch can | | reorder these two writes | | loop long between WRITE_ONCE(node->rb_left, parent);-+-+ parent and node | +--->finally see parent->rb_right
Quoted 4 lines
Quoted 2 lines
The long loop won't stop until the modifer's CPU flushes its writes. Too avoid it, we should limit the searching depth.
Cute, have you actually observed this? Did you have performance issues?
I can only test it on x86 by now, which implies smp_wmb() between writes. I haven't observed any thing wrong. I'm just imaging it on some other ARCHs.
I accidentally found this part of code when I searched for whether there is any attempt again to use rbtree with RCU, and whether there are the cases besides speculative page fault.
Quoted 11 lines
Quoted 7 lines
There are no more than (1<<BITS_PER_LONG)-1 nodes in the tree. And the
max_depthof a tree is no more than 2*lg(node_count+1), which is no mare than 2*BITS_PER_LONG.So the serarch should stop when diving down up to
2*BITS_PER_LONG depth.Arguably you can have a larger key space, but I think due to memory constraints this limit still isn't wrong. But I do feel you need a comment with that.
Sure, I will add some comments about why "2*BITS_PER_LONG" in code.
But how it could be larger key space? there are not more than (1<<BITS_PER_LONG) bytes in the kernel dereferencable address space, and (1<<BITS_PER_LONG)/sizeof(rb_node) must be less than (1<<BITS_PER_LONG)-1.
Quoted 23 lines · Lai Jiangshan
Quoted 22 lines · Peter Zijlstra
Quoted 21 lines · Lai Jiangshan
lib/rbtree.chas ensured that there is not possible to inadvertently cause (temporary) loops in the tree structure as seen in program order of the modifier. But loop is still possible to be seen in searcher due to CPU's reordering.for example:
modifier searcherleft rotate at parent
parent->rb_rightis nodesearch to parent parent->rb_right is node +->see node->rb_left changed WRITE_ONCE(parent->rb_right, tmp);-+ | node->rb_left is parennt no smp_wmb(), some arch can | | reorder these two writes | | loop long between WRITE_ONCE(node->rb_left, parent);-+-+ parent and node | +--->finally see parent->rb_right
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
The long loop won't stop until the modifer's CPU flushes its writes. Too avoid it, we should limit the searching depth.
Cute, have you actually observed this? Did you have performance issues?
I can only test it on x86 by now, which implies
smp_wmb()between writes. I haven't observed any thing wrong. I'm just imaging it on some other ARCHs.
Note that smp_wmb() doesn't imply flushing the store-buffers. Nor does the TSO memory model of x86 (it's the main feature that distinguishes TSO from SC).
x86's MFENCE is a completion barrier and does imply so though.
Quoted 3 lines
I accidentally found this part of code when I searched for whether there is any attempt again to use rbtree with RCU, and whether there are the cases besides speculative page fault.
It got mentioned earlier in the contect of a stream of changes, an uninterrupted modifier can basically starve a search.
But I don't think that's a problem with the current users.
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines
There are no more than (1<<BITS_PER_LONG)-1 nodes in the tree. And the
max_depthof a tree is no more than 2*lg(node_count+1), which is no mare than 2*BITS_PER_LONG.So the serarch should stop when diving down up to
2*BITS_PER_LONG depth.Arguably you can have a larger key space, but I think due to memory constraints this limit still isn't wrong. But I do feel you need a comment with that.
Quoted 6 lines
Sure, I will add some comments about why "2*BITS_PER_LONG" in code.
But how it could be larger key space? there are not more than (1<<BITS_PER_LONG) bytes in the kernel dereferencable address space, and (1<<BITS_PER_LONG)/sizeof(rb_node) must be less than (1<<BITS_PER_LONG)-1.
Well, the key space is determined by the comparator operators that are provided, which can easily compare values that are larger than 64bit.
But yes, the address space implies limits regardless of the actual key-space. Note that BITS_PER_LONG does not related to the actual memory space for things like i386-PAE and ARMv7-LPEA.
PATCH v2 1/2 rbtree_latch: quit searching when reaching to maximum depth
lib/rbtree.c has ensured that there is not possible to inadvertently cause (temporary) loops in the tree structure as seen in program order of the modifier. But loop is still possible to be seen in searcher due to CPU's reordering.
for example:
modifier searcher
left rotate at parentparent->rb_right is node
search to parent parent->rb_right is node +->see node->rb_left changed WRITE_ONCE(parent->rb_right, tmp);-+ | node->rb_left is parennt no smp_wmb(), some ARCHs can | | reorder these two writes | | loop long between WRITE_ONCE(node->rb_left, parent);-+-+ parent and node | +--->finally see parent->rb_right
The long loop won't stop until the modifer's CPU flushes its writes. Too avoid it, we should limit the searching depth. There are no more than (1<<BITS_PER_LONG)-1 nodes in the tree. And the max_depth of a tree is no more than 2*lg(node_count+1), which is no mare than 2*BITS_PER_LONG.
So the searching should stop when diving down up to 2*BITS_PER_LONG depth.
--- include/linux/rbtree_latch.h | 39 ++++++++++++++++++++++++++++++++++++ 1 file changed, 39 insertions(+)
include/linux/rbtree_latch.h +39 -0
@@ -102,11 +102,47 @@ __lt_erase(struct latch_tree_node *ltn, struct latch_tree_root *ltr, int idx)rb_erase (<n->node[idx], <r->tree[idx]); } +/* +* Beware when rbtree is being searched in RCU read sites. +* +* lib/rbtree.c has ensured that there is not possible to +* inadvertently cause (temporary) loops in the tree structure +* as seen in program order of the modifier. But loop is still +* possible to be seen in searcher due to CPU's reordering. +* +* for example: +* modifier searcher +* +* left rotate at parent search to parent +* parent->rb_right is node parent->rb_right is node +* +->see node->rb_left changed +* WRITE_ONCE(parent->rb_right, tmp);-+ | node->rb_left is parennt +* no smp_wmb(), some ARCHs can | | +* reorder these two writes | | loop long between +* WRITE_ONCE(node->rb_left, parent);-+-+ parent and node +* | +* +--->finally see +* parent->rb_right +* +* The long loop won't stop until the searcher finally see the changes +* from the modifier. Too avoid it, we should limit the searching depth. +* +* Limited by the address space of the kernel, there are no more than +* (1<<BITS_PER_LONG)-1 nodes in the tree. And the max_depth of a tree is +* no more than 2*lg(node_count+1), which is no mare than 2*BITS_PER_LONG. +* +* So the searching should stop when diving down up to +* 2*BITS_PER_LONG depth. +* +* Note: the above problem is not subject to the TSO memory model, such as +* x86, which can dispense with the depth check. +*/ static __always_inline struct latch_tree_node *__lt_find (void *key,struct latch_tree_root *ltr,int idx,int (*comp)(void *key,struct latch_tree_node *node)) {struct rb_node *node =rcu_dereference_raw (ltr->tree[idx].rb_node); +int depth =2 *BITS_PER_LONG ;struct latch_tree_node *ltn;int c;
@@ -120,6 +156,9 @@ __lt_find(void *key, struct latch_tree_root *ltr, int idx, node =rcu_dereference_raw (node->rb_right); else return ltn; + +if (!IS_ENABLED (CONFIG_X86 ) && (--depth <0 )) +break ; }return NULL ;
Signature · 1 line
2.20.1
PATCH v2 2/2 rbtree_latch: don't need to check seq when it found a node
latch_tree_find() should be protected by caller via RCU or so. When it find a node in an attempt, the node must be a valid one in RCU's point's of view even the tree is (being) updated with a new node with the same key which is entirely subject to timing anyway.
--- include/linux/rbtree_latch.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
include/linux/rbtree_latch.h +1 -1
@@ -245,7 +245,7 @@ latch_tree_find(void *key, struct latch_tree_root *root,do { seq =raw_read_seqcount_latch (&root->seq); node =__lt_find (key, root, seq &1 , ops->comp); - }while (read_seqcount_retry (&root->seq, seq)); + }while (!node &&read_seqcount_retry (&root->seq, seq));return node; }
Signature · 1 line
2.20.1
Re: [PATCH V2 2/2] rbtree_latch: don't need to check seq when it found a node
Quoted 5 lines · Lai Jiangshan
latch_tree_find()should be protected by caller via RCU or so. When it find a node in an attempt, the node must be a valid one in RCU's point's of view even the tree is (being) updated with a new node with the same key which is entirely subject to timing anyway.
I'm not sure I buy this. Even if we get a valid node, is it the one we were searching for ? I don't see how this could be guaranteed if the read raced with a tree rebalancing.
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Re: [PATCH V2 2/2] rbtree_latch: don't need to check seq when it found a node
Quoted 11 lines · Michel Lespinasse
Quoted 5 lines · Lai Jiangshan
latch_tree_find()should be protected by caller via RCU or so. When it find a node in an attempt, the node must be a valid one in RCU's point's of view even the tree is (being) updated with a new node with the same key which is entirely subject to timing anyway.I'm not sure I buy this. Even if we get a valid node, is it the one we were searching for ? I don't see how this could be guaranteed if the read raced with a tree rebalancing.
It is valid because ops->comp() returns 0 and it should be the one we were searching for unless ops->comp() is wrong. The searched one could be possible just deleted, but it is still a legitimate searched result in RCU's point's of view.
A tree rebalancing can cause a searching fails to find an existing target. This is the job of read_seqcount_retry() to tell you to retry.
Quoted 4 lines
--
Michel "Walken" Lespinasse
A program is never fully debugged until the last user dies.
Re: [PATCH V2 2/2] rbtree_latch: don't need to check seq when it found a node
Quoted 14 lines · Lai Jiangshan
Quoted 11 lines · Michel Lespinasse
Quoted 5 lines · Lai Jiangshan
latch_tree_find()should be protected by caller via RCU or so. When it find a node in an attempt, the node must be a valid one in RCU's point's of view even the tree is (being) updated with a new node with the same key which is entirely subject to timing anyway.I'm not sure I buy this. Even if we get a valid node, is it the one we were searching for ? I don't see how this could be guaranteed if the read raced with a tree rebalancing.
Quoted 8 lines
It is valid because
ops->comp()returns 0 and it should be the one we were searching for unlessops->comp()is wrong. The searched one could be possible just deleted, but it is still a legitimate searched result in RCU's point's of view.A tree rebalancing can cause a searching fails to find an existing target. This is the job of
read_seqcount_retry()to tell you to retry.
Ah, yes, this is correct. It wouldn't work if we wanted to return the next higher key for example, but it does work for exact matches. Nice!
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Quoted 39 lines · Lai Jiangshan
lib/rbtree.chas ensured that there is not possible to inadvertently cause (temporary) loops in the tree structure as seen in program order of the modifier. But loop is still possible to be seen in searcher due to CPU's reordering.for example:
modifier searcherleft rotate at parent
parent->rb_rightis nodesearch to parent parent->rb_right is node +->see node->rb_left changed WRITE_ONCE(parent->rb_right, tmp);-+ | node->rb_left is parennt no smp_wmb(), some ARCHs can | | reorder these two writes | | loop long between WRITE_ONCE(node->rb_left, parent);-+-+ parent and node | +--->finally see parent->rb_rightThe long loop won't stop until the modifer's CPU flushes its writes. Too avoid it, we should limit the searching depth. There are no more than (1<<BITS_PER_LONG)-1 nodes in the tree. And the
max_depthof a tree is no more than 2*lg(node_count+1), which is no mare than 2*BITS_PER_LONG.So the searching should stop when diving down up to 2*BITS_PER_LONG depth.
Cc Peter Zijlstra <[email protected]>Cc Paul E. McKenney <[email protected]>Cc Oleg Nesterov <[email protected]>Cc Michel Lespinasse <[email protected]>Cc Andrea Arcangeli <[email protected]>Cc Rik van Riel <[email protected]>Cc Mathieu Desnoyers <[email protected]>Signed-off-by Lai Jiangshan <[email protected]>
Quoted 1 line
---
Hello
Could anyone have a review or an ack on the updated patch? Compared v1:
Applied Mathieu's suggest to change the changelog Avoid the depth check on x86, so the patch makes no functionality
change on x86
And which tree should the patches route to? It is memory ordering related.
Thanks
Lai
Quoted 68 lines
include/linux/rbtree_latch.h | 39 ++++++++++++++++++++++++++++++++++++ 1 file changed, 39 insertions(+)include/linux/rbtree_latch.h +39
-1@@ -102,11 +102,47 @@ __lt_erase(struct latch_tree_node *ltn, struct latch_tree_root *ltr, int idx)rb_erase (<n->node[idx], <r->tree[idx]); } +/* +* Beware when rbtree is being searched in RCU read sites. +* +* lib/rbtree.c has ensured that there is not possible to +* inadvertently cause (temporary) loops in the tree structure +* as seen in program order of the modifier. But loop is still +* possible to be seen in searcher due to CPU's reordering. +* +* for example: +* modifier searcher +* +* left rotate at parent search to parent +* parent->rb_right is node parent->rb_right is node +* +->see node->rb_left changed +* WRITE_ONCE(parent->rb_right, tmp);-+ | node->rb_left is parennt +* no smp_wmb(), some ARCHs can | | +* reorder these two writes | | loop long between +* WRITE_ONCE(node->rb_left, parent);-+-+ parent and node +* | +* +--->finally see +* parent->rb_right +* +* The long loop won't stop until the searcher finally see the changes +* from the modifier. Too avoid it, we should limit the searching depth. +* +* Limited by the address space of the kernel, there are no more than +* (1<<BITS_PER_LONG)-1 nodes in the tree. And the max_depth of a tree is +* no more than 2*lg(node_count+1), which is no mare than 2*BITS_PER_LONG. +* +* So the searching should stop when diving down up to +* 2*BITS_PER_LONG depth. +* +* Note: the above problem is not subject to the TSO memory model, such as +* x86, which can dispense with the depth check. +*/ static __always_inline struct latch_tree_node *__lt_find (void *key,struct latch_tree_root *ltr,int idx,int (*comp)(void *key,struct latch_tree_node *node)) {struct rb_node *node =rcu_dereference_raw (ltr->tree[idx].rb_node); +int depth =2 *BITS_PER_LONG ;struct latch_tree_node *ltn;int c;@@ -120,6 +156,9 @@ __lt_find(void *key, struct latch_tree_root *ltr, int idx, node =rcu_dereference_raw (node->rb_right); else return ltn; + +if (!IS_ENABLED (CONFIG_X86 ) && (--depth <0 )) +break ; }return NULL ; --2.20.1
----- On May 15, 2020, at 8:47 AM, Lai Jiangshan [email protected] wrote:
[...]
Quoted 2 lines
So the serarch should stop when diving down up to
2*BITS_PER_LONG depth.
serarch -> search
Thanks,
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com