Re: [PATCH bpf-next] libbpf: don't allocate 16M for log buffer by default

On 3/25/20 1:08 AM, Andrii Nakryiko wrote:

On Tue, Mar 24, 2020 at 4:59 PM Stanislav Fomichev [off-list ref] wrote:

quoted

On 03/24, Andrii Nakryiko wrote:

quoted

On Tue, Mar 24, 2020 at 4:31 PM Stanislav Fomichev [off-list ref] wrote:

quoted

For each prog/btf load we allocate and free 16 megs of verifier buffer.
On production systems it doesn't really make sense because the
programs/btf have gone through extensive testing and (mostly) guaranteed
to successfully load.

Let's switch to a much smaller buffer by default (128 bytes, sys_bpf
doesn't accept smaller log buffer) and resize it if the kernel returns
ENOSPC. On the first ENOSPC error we resize the buffer to BPF_LOG_BUF_SIZE
and then, on each subsequent ENOSPC, we keep doubling the buffer.

Signed-off-by: Stanislav Fomichev <redacted>
---
  tools/lib/bpf/btf.c             | 10 +++++++++-
  tools/lib/bpf/libbpf.c          | 10 ++++++++--
  tools/lib/bpf/libbpf_internal.h |  2 ++
  3 files changed, 19 insertions(+), 3 deletions(-)

diff --git a/tools/lib/bpf/btf.c b/tools/lib/bpf/btf.c
index 3d1c25fc97ae..53c7efc3b347 100644
--- a/tools/lib/bpf/btf.c
+++ b/tools/lib/bpf/btf.c

@@ -657,13 +657,14 @@ int btf__finalize_data(struct bpf_object *obj, struct btf *btf)

  int btf__load(struct btf *btf)
  {
-       __u32 log_buf_size = BPF_LOG_BUF_SIZE;
+       __u32 log_buf_size = BPF_MIN_LOG_BUF_SIZE;
         char *log_buf = NULL;
         int err = 0;

         if (btf->fd >= 0)
                 return -EEXIST;

+retry_load:
         log_buf = malloc(log_buf_size);
         if (!log_buf)
                 return -ENOMEM;

I'd argue that on first try we shouldn't allocate log_buf at all, then
start allocating it using reasonable starting size (see below).

Agreed, makes sense.

The iproute2 BPF loader does the first try without any log buffer, and then
successively increases the size on failure [0].

libbpf should also assume the success case in the very first run, and only
then redo the load attempt with log buffer to avoid the overhead.

   [0] https://git.kernel.org/pub/scm/network/iproute2/iproute2.git/tree/lib/bpf.c

quoted

quoted

quoted

@@ -673,6 +674,13 @@ int btf__load(struct btf *btf)
         btf->fd = bpf_load_btf(btf->data, btf->data_size,
                                log_buf, log_buf_size, false);
         if (btf->fd < 0) {
+               if (errno == ENOSPC) {
+                       log_buf_size = max((__u32)BPF_LOG_BUF_SIZE,
+                                          log_buf_size << 1);
+                       free(log_buf);
+                       goto retry_load;
+               }
+
                 err = -errno;
                 pr_warn("Error loading BTF: %s(%d)\n", strerror(errno), errno);
                 if (*log_buf)

[...]

quoted

quoted

quoted

diff --git a/tools/lib/bpf/libbpf_internal.h b/tools/lib/bpf/libbpf_internal.h
index 8c3afbd97747..2720f3366798 100644
--- a/tools/lib/bpf/libbpf_internal.h
+++ b/tools/lib/bpf/libbpf_internal.h

@@ -23,6 +23,8 @@
  #define BTF_PARAM_ENC(name, type) (name), (type)
  #define BTF_VAR_SECINFO_ENC(type, offset, size) (type), (offset), (size)

+#define BPF_MIN_LOG_BUF_SIZE 128

This seems way too low, if there is some error it almost certainly
will be too short, probably for few iterations, just causing waste.
Let's make it something a bit more reasonable, like 32KB or something?

In this case, maybe start with the existing 16M BPF_LOG_BUF_SIZE?
My goal here is optimize for the successful case. If there is an error the
size shouldn't matter that much.

Not feeling strongly. But we already will have a retry loop, so not
too hard to do it in steps. Then also errors do happen in production
as well, and it would be good to not eat too much memory
unnecessarily.

Yeah, 128 is way too low. Either we should just malloc the max possible size
right away, or make it a two-step approach where we first try with half the
size and only if that fails we retry with the max size. Given programs can be
very complex this will otherwise just prolong the time unnecessarily for the
failure verdict and unnecessarily puts load on the verifier.

Thanks,
Daniel