Skip to contentSkip to conversation

bpf: use kvmalloc in map_lookup_elem

PATCH In review Stanislav Fomichev 3 participants 4 messages Aug 16, 2021

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bpf: use kvmalloc in map_lookup_elem

Use kvmalloc/kvfree for temporary value when looking up a map. kmalloc might not be sufficient for percpu maps where the value is big.

Can be reproduced with netcnt test on qemu with "-smp 255".

---
 kernel/bpf/syscall.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
kernel/bpf/syscall.c +2 -2
index 9a2068e39d23..ae0b1c1c8ece 100644
@@ -1076,7 +1076,7 @@ static int map_lookup_elem(union bpf_attr *attr)
 	value_size = bpf_map_value_size(map);
 
 	err = -ENOMEM;
-	value = kmalloc(value_size, GFP_USER | __GFP_NOWARN);
+	value = kvmalloc(value_size, GFP_USER | __GFP_NOWARN);
 	if (!value)
 		goto free_key;
 
@@ -1091,7 +1091,7 @@ static int map_lookup_elem(union bpf_attr *attr)
 	err = 0;
 
 free_value:
-	kfree(value);
+	kvfree(value);
 free_key:
 	kfree(key);
 err_put:
Signature · 1 line
2.33.0.rc1.237.g0d66db33f3-goog
maintainer
#1 lore raw
Quoted 15 lines · Stanislav Fomichev

Use kvmalloc/kvfree for temporary value when looking up a map. kmalloc might not be sufficient for percpu maps where the value is big.

Can be reproduced with netcnt test on qemu with "-smp 255".

---
  kernel/bpf/syscall.c | 4 ++--
  1 file changed, 2 insertions(+), 2 deletions(-)
kernel/bpf/syscall.c +0 -0
index 9a2068e39d23..ae0b1c1c8ece 100644

kernel/bpf/syscall.c static int map_lookup_elem(union bpf_attr *attr) · line 1081

Quoted 7 lines
kernel/bpf/syscall.c +1 -1
@@ -1076,7 +1076,7 @@ static int map_lookup_elem(union bpf_attr *attr)
  	value_size = bpf_map_value_size(map);
 
  	err = -ENOMEM;
-	value = kmalloc(value_size, GFP_USER | __GFP_NOWARN);
+	value = kvmalloc(value_size, GFP_USER | __GFP_NOWARN);
  	if (!value)
  		goto free_key;

What about other cases like map_update_elem(), shouldn't they be adapted similarly?

Thanks,
Daniel

#2 lore raw
Quoted 19 lines · Daniel Borkmann
Quoted 17 lines · Stanislav Fomichev

Use kvmalloc/kvfree for temporary value when looking up a map. kmalloc might not be sufficient for percpu maps where the value is big.

Can be reproduced with netcnt test on qemu with "-smp 255".

---
  kernel/bpf/syscall.c | 4 ++--
  1 file changed, 2 insertions(+), 2 deletions(-)
kernel/bpf/syscall.c +0 -0
index 9a2068e39d23..ae0b1c1c8ece 100644
@@ -1076,7 +1076,7 @@ static int map_lookup_elem(union bpf_attr *attr)
      value_size = bpf_map_value_size(map);
 
Quoted 8 lines

kernel/bpf/syscall.c static int map_lookup_elem(union bpf_attr *attr) · line 1081

Quoted 5 lines
kernel/bpf/syscall.c +1 -1
      err = -ENOMEM;
-     value = kmalloc(value_size, GFP_USER | __GFP_NOWARN);
+     value = kvmalloc(value_size, GFP_USER | __GFP_NOWARN);
      if (!value)
              goto free_key;

What about other cases like map_update_elem(), shouldn't they be adapted similarly?

And in the same vein (with keys potentially being big as well), should we switch __bpf_copy_key() to use vmemdup_user() instead of memdup_user()?

Quoted 3 lines

Thanks,
Daniel

#3 lore raw
Quoted 30 lines · Andrii Nakryiko
Quoted 4 lines · Daniel Borkmann
Quoted 2 lines · Stanislav Fomichev

Use kvmalloc/kvfree for temporary value when looking up a map. kmalloc might not be sufficient for percpu maps where the value is

big.

Quoted 15 lines
Quoted 15 lines

Can be reproduced with netcnt test on qemu with "-smp 255".

---
  kernel/bpf/syscall.c | 4 ++--
  1 file changed, 2 insertions(+), 2 deletions(-)
kernel/bpf/syscall.c +0 -0
index 9a2068e39d23..ae0b1c1c8ece 100644
@@ -1076,7 +1076,7 @@ static int map_lookup_elem(union bpf_attr *attr)
      value_size = bpf_map_value_size(map);
 
Quoted 8 lines

kernel/bpf/syscall.c static int map_lookup_elem(union bpf_attr *attr) · line 1081

Quoted 5 lines
kernel/bpf/syscall.c +1 -1
      err = -ENOMEM;
-     value = kmalloc(value_size, GFP_USER | __GFP_NOWARN);
+     value = kvmalloc(value_size, GFP_USER | __GFP_NOWARN);
      if (!value)
              goto free_key;

What about other cases like map_update_elem(), shouldn't they be adapted similarly?

Quoted 3 lines

And in the same vein (with keys potentially being big as well), should we switch __bpf_copy_key() to use vmemdup_user() instead of memdup_user()?

Those are good questions :-)

I'm assuming that whatever is doing kmalloc on top of bpf_map_value_size() should definitely use kvmalloc since bpf_map_value_size() can blow up the value size. (will resend)

For __bpf_copy_key I'm less sure, but it might be a good idea as well. Let me try to look at bit more into it, but feels like there shouldn't be any downsides?