Skip to contentSkip to conversation

bpf: Use kmalloc_nolock() in range tree

PATCH Superseded Puranjay Mohan 3 participants 3 messages Nov 6, 2025 BPF

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bpf: Use kmalloc_nolock() in range tree

The range tree uses bpf_mem_alloc() that is safe to be called from all contexts and uses a pre-allocated pool of memory to serve these allocations.

Replace bpf_mem_alloc() with kmalloc_nolock() as it can be called safely from all contexts and is more scalable than bpf_mem_alloc().

Remove the migrate_disable/enable pairs as they were only needed for bpf_mem_alloc() as it does per-cpu operations, kmalloc_nolock() doesn't need this.

---
 kernel/bpf/range_tree.c | 22 +++++++---------------
 1 file changed, 7 insertions(+), 15 deletions(-)
kernel/bpf/range_tree.c +7 -15
index 37b80a23ae1a..2f28886f3ff7 100644
@@ -2,7 +2,6 @@
 /* Copyright (c) 2024 Meta Platforms, Inc. and affiliates. */
 #include <linux/interval_tree_generic.h>
 #include <linux/slab.h>
-#include <linux/bpf_mem_alloc.h>
 #include <linux/bpf.h>
 #include "range_tree.h"
 
@@ -21,7 +20,7 @@
  * in commit 6772fcc8890a ("xfs: convert xbitmap to interval tree").
  *
  * The implementation relies on external lock to protect rbtree-s.
- * The alloc/free of range_node-s is done via bpf_mem_alloc.
+ * The alloc/free of range_node-s is done via kmalloc_nolock().
  *
  * bpf arena is using range_tree to represent unallocated slots.
  * At init time:
@@ -150,9 +149,8 @@ int range_tree_clear(struct range_tree *rt, u32 start, u32 len)
 			range_it_insert(rn, rt);
 
 			/* Add a range */
-			migrate_disable();
-			new_rn = bpf_mem_alloc(&bpf_global_ma, sizeof(struct range_node));
-			migrate_enable();
+			new_rn = kmalloc_nolock(sizeof(struct range_node), __GFP_ACCOUNT,
+						NUMA_NO_NODE);
 			if (!new_rn)
 				return -ENOMEM;
 			new_rn->rn_start = last + 1;
@@ -172,9 +170,7 @@ int range_tree_clear(struct range_tree *rt, u32 start, u32 len)
 		} else {
 			/* in the middle of the clearing range */
 			range_it_remove(rn, rt);
-			migrate_disable();
-			bpf_mem_free(&bpf_global_ma, rn);
-			migrate_enable();
+			kfree_nolock(rn);
 		}
 	}
 	return 0;
@@ -227,9 +223,7 @@ int range_tree_set(struct range_tree *rt, u32 start, u32 len)
 		range_it_remove(right, rt);
 		left->rn_last = right->rn_last;
 		range_it_insert(left, rt);
-		migrate_disable();
-		bpf_mem_free(&bpf_global_ma, right);
-		migrate_enable();
+		kfree_nolock(right);
 	} else if (left) {
 		/* Combine with the left range */
 		range_it_remove(left, rt);
@@ -241,9 +235,7 @@ int range_tree_set(struct range_tree *rt, u32 start, u32 len)
 		right->rn_start = start;
 		range_it_insert(right, rt);
 	} else {
-		migrate_disable();
-		left = bpf_mem_alloc(&bpf_global_ma, sizeof(struct range_node));
-		migrate_enable();
+		left = kmalloc_nolock(sizeof(struct range_node), __GFP_ACCOUNT, NUMA_NO_NODE);
 		if (!left)
 			return -ENOMEM;
 		left->rn_start = start;
@@ -259,7 +251,7 @@ void range_tree_destroy(struct range_tree *rt)
 
 	while ((rn = range_it_iter_first(rt, 0, -1U))) {
 		range_it_remove(rn, rt);
-		bpf_mem_free(&bpf_global_ma, rn);
+		kfree_nolock(rn);
 	}
 }
 
Signature · 1 line
2.47.3
#1 lore raw
Quoted 39 lines · Puranjay Mohan

The range tree uses bpf_mem_alloc() that is safe to be called from all contexts and uses a pre-allocated pool of memory to serve these allocations.

Replace bpf_mem_alloc() with kmalloc_nolock() as it can be called safely from all contexts and is more scalable than bpf_mem_alloc().

Remove the migrate_disable/enable pairs as they were only needed for bpf_mem_alloc() as it does per-cpu operations, kmalloc_nolock() doesn't need this.

---
 kernel/bpf/range_tree.c | 22 +++++++---------------
 1 file changed, 7 insertions(+), 15 deletions(-)
kernel/bpf/range_tree.c +1 -2
index 37b80a23ae1a..2f28886f3ff7 100644
@@ -2,7 +2,6 @@
 /* Copyright (c) 2024 Meta Platforms, Inc. and affiliates. */
 #include <linux/interval_tree_generic.h>
 #include <linux/slab.h>
-#include <linux/bpf_mem_alloc.h>
 #include <linux/bpf.h>
 #include "range_tree.h"
 
@@ -21,7 +20,7 @@
  * in commit 6772fcc8890a ("xfs: convert xbitmap to interval tree").
  *
  * The implementation relies on external lock to protect rbtree-s.
- * The alloc/free of range_node-s is done via bpf_mem_alloc.
+ * The alloc/free of range_node-s is done via kmalloc_nolock().
  *
  * bpf arena is using range_tree to represent unallocated slots.
  * At init time:

kernel/bpf/range_tree.c int range_tree_clear(struct range_tree *rt, u32 start, u32 len) · line 153

Quoted 8 lines
kernel/bpf/range_tree.c +2 -3
@@ -150,9 +149,8 @@ int range_tree_clear(struct range_tree *rt, u32 start, u32 len)
                        range_it_insert(rn, rt);
 
                        /* Add a range */
-                       migrate_disable();
-                       new_rn = bpf_mem_alloc(&bpf_global_ma, sizeof(struct range_node));
-                       migrate_enable();
+                       new_rn = kmalloc_nolock(sizeof(struct range_node), __GFP_ACCOUNT,
+                                               NUMA_NO_NODE);

bpf_global_ma would consistently charge root memcg, since it is saved at init time, while above kmalloc_nolock() will charge random current task.
Let's drop __GFP_ACCOUNT.
The rest looks good.

pw-bot: cr

#2 lore raw
Quoted 43 lines · Alexei Starovoitov
Quoted 41 lines · Puranjay Mohan

The range tree uses bpf_mem_alloc() that is safe to be called from all contexts and uses a pre-allocated pool of memory to serve these allocations.

Replace bpf_mem_alloc() with kmalloc_nolock() as it can be called safely from all contexts and is more scalable than bpf_mem_alloc().

Remove the migrate_disable/enable pairs as they were only needed for bpf_mem_alloc() as it does per-cpu operations, kmalloc_nolock() doesn't need this.

---
 kernel/bpf/range_tree.c | 22 +++++++---------------
 1 file changed, 7 insertions(+), 15 deletions(-)
kernel/bpf/range_tree.c +1 -2
index 37b80a23ae1a..2f28886f3ff7 100644
@@ -2,7 +2,6 @@
 /* Copyright (c) 2024 Meta Platforms, Inc. and affiliates. */
 #include <linux/interval_tree_generic.h>
 #include <linux/slab.h>
-#include <linux/bpf_mem_alloc.h>
 #include <linux/bpf.h>
 #include "range_tree.h"
 
@@ -21,7 +20,7 @@
  * in commit 6772fcc8890a ("xfs: convert xbitmap to interval tree").
  *
  * The implementation relies on external lock to protect rbtree-s.
- * The alloc/free of range_node-s is done via bpf_mem_alloc.
+ * The alloc/free of range_node-s is done via kmalloc_nolock().
  *
  * bpf arena is using range_tree to represent unallocated slots.
  * At init time:
@@ -150,9 +149,8 @@ int range_tree_clear(struct range_tree *rt, u32 start, u32 len)
                        range_it_insert(rn, rt);
 
Quoted 12 lines

kernel/bpf/range_tree.c int range_tree_clear(struct range_tree *rt, u32 start, u32 len) · line 153

Quoted 6 lines
kernel/bpf/range_tree.c +2 -3
                        /* Add a range */
-                       migrate_disable();
-                       new_rn = bpf_mem_alloc(&bpf_global_ma, sizeof(struct range_node));
-                       migrate_enable();
+                       new_rn = kmalloc_nolock(sizeof(struct range_node), __GFP_ACCOUNT,
+                                               NUMA_NO_NODE);

bpf_global_ma would consistently charge root memcg, since it is saved at init time, while above kmalloc_nolock() will charge random current task.
Let's drop __GFP_ACCOUNT.
The rest looks good.

Sure,
as this change is small I will just send v2 right away.