On Fri, Jul 12, 2024 at 6:50 AM Daniel Borkmann <daniel@xxxxxxxxxxxxx> wrote: > > Hi Puranjay, > > On 7/11/24 4:00 PM, Puranjay Mohan wrote: > [...] > > I was able find the root cause of this bug and will send a fix soon! > > > >> Unable to handle kernel paging request at virtual address ffff0000c2a80e68 > > > > We are running this test on Qemu with '-cpu max', this means 52-bit > > virtual addresses are being used. > > > > The trampolines generation code has the following two lines: > > > > emit_addr_mov_i64(A64_R(0), (const u64)im, ctx); > > emit_call((const u64)__bpf_tramp_enter, ctx); > > > > here the address of struct bpf_tramp_image is moved to R0 and passed as > > an argument to __bpf_tramp_enter(). > > > > emit_addr_mov_i64() assumes that the address passed to it is in the > > vmalloc space and uses at most 48 bits. It sets all the remaining bits > > to 1. > > > > but struct bpf_tramp_image is allocated using kzalloc() and when 52-bit > > VAs are used, its address is not guaranteed to be 48-bit, therefore we > > see this bug, where 0xfff[0]0000c2a80e68 is converted to > > 0xfff[f]0000c2a80e68 when the trampoline is generated. > > > > The fix would be use emit_a64_mov_i64() for moving this address into R0. > > It looks like there is still an issue left. A recent CI run on bpf-next is > still hitting the same on arm64: > > Base: > > https://github.com/kernel-patches/bpf/commits/series/870746%3D%3Ebpf-next/ > > CI: > > https://github.com/kernel-patches/bpf/actions/runs/9905842936/job/27366435436 > > [...] > #89/11 fexit_bpf2bpf/func_replace_global_func:OK > #89/12 fexit_bpf2bpf/fentry_to_cgroup_bpf:OK > #89/13 fexit_bpf2bpf/func_replace_progmap:OK > #89 fexit_bpf2bpf:OK > Error: The operation was canceled. Let's denylist that test again for now?
