On Tue, Oct 8, 2024 at 3:10 PM Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > > We need to scrap this idea. > Let's go back to push/pop r11 around calls :( I didn't give up :) Here is a new idea that seems to work: [ 131.472066] dump_stack_lvl+0x53/0x70 [ 131.472066] bpf_task_storage_get+0x3e/0x2f0 [ 131.472066] ? bpf_task_storage_get+0x231/0x2f0 [ 131.472066] bpf_prog_ed7a5f33cc9fefab_foo+0x30/0x32 [ 131.472066] bpf_prog_8c4f9bc79da6c27e_socket_post_create+0x68/0x6d ... [ 131.417145] dump_stack_lvl+0x53/0x70 [ 131.417145] bpf_task_storage_get+0x3e/0x2f0 [ 131.417145] ? selinux_netlbl_socket_post_create+0xab/0x150 [ 131.417145] bpf_prog_8c4f9bc79da6c27e_socket_post_create+0x60/0x6d The stack dump works fine out of main prog and out of subprog. The key difference it to pretend to have stack_depth=0, so there is no adjustment to %rsp, but point %rbp to per-cpu private stack and grow it _up_. For the main prog %rbp points to the bottom of priv stack plus stack_depth it needs, so all bpf insns that do r10-off access the bottom of that priv stack. When subprog is called it does 'add %rbp, its_stack_depth' and in turn it's using memory above the bottom of the priv stack. That seems to work, but exceptions and tailcalls are broken. I ran out of time today to debug. Pls see the attached patch.
Attachment:
0001-bpf-Private-stack-via-rbp.patch
Description: Binary data
