Re: [PATCH bpf-next v3 06/10] bpf: Bump MAX_BPF_STACK size to 768 bytes

On Wed, Sep 15, 2021 at 10:57 AM Kumar Kartikeya Dwivedi
[off-list ref] wrote:

On Wed, Sep 15, 2021 at 10:03:53PM IST, Alexei Starovoitov wrote:

quoted

On Wed, Sep 15, 2021 at 10:39:39AM +0530, Kumar Kartikeya Dwivedi wrote:

quoted

Increase the maximum stack size accessible to BPF program to 768 bytes.
This is done so that gen_loader can use 94 additional fds for kfunc BTFs
that it passes in to fd_array from the remaining space available for the
loader_stack struct to expand.

Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
---
 include/linux/filter.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/linux/filter.h b/include/linux/filter.h
index 4a93c12543ee..b214189ece62 100644
--- a/include/linux/filter.h
+++ b/include/linux/filter.h

@@ -82,8 +82,8 @@ struct ctl_table_header;
  */
 #define BPF_SYM_ELF_TYPE   't'

-/* BPF program can access up to 512 bytes of stack space. */
-#define MAX_BPF_STACK      512
+/* BPF program can access up to 768 bytes of stack space. */
+#define MAX_BPF_STACK      768

Yikes.
I guess you meant as RFC, right? You didn't really propose
to increase prog stack size just for that, right?

Yes, and right, it's ugly :/.

quoted

In the later patch:
+/* MAX_BPF_STACK is 768 bytes, so (64 + 32 + 94 (MAX_KFUNC_DESCS) + 2) * 4 */
 #define MAX_USED_MAPS 64
 #define MAX_USED_PROGS 32

@@ -31,6 +33,8 @@ struct loader_stack {
        __u32 btf_fd;
        __u32 map_fd[MAX_USED_MAPS];
        __u32 prog_fd[MAX_USED_PROGS];
+       /* Update insn->off store when reordering kfunc_btf_fd */
+       __u32 kfunc_btf_fd[MAX_KFUNC_DESCS];
        __u32 inner_map_fd;

};

There are few other ways to do that.
For example:
A: rename map_fd[] into fds[] and store both map and btf FDs in there.
B: move map and btf FDs into data instead of stack.

Both are great suggestions, I thought about A but not B, but it will be better
(even though it requires more changes, we can do full 256 BTF fds using B).
Thanks!

btw fd_array doesn't have to have valid FDs in all slots
when passed into the prog_load command.
If bpf prog doesn't index into them they can have garbage.
Just mentioning in case that simplifies the implementation.
I suspect packing btf_fds right after map_fds with no gaps will not be
hard to do, but in case it's somehow too painful there could be a gap.