Björn Töpel <bjorn@xxxxxxxxxx> writes:
> Palmer Dabbelt <palmer@xxxxxxxxxxxx> writes:
>
>> On Wed, 12 Jul 2023 11:11:08 PDT (-0700), bjorn@xxxxxxxxxx wrote:
>>> Song Shuai <suagrfillet@xxxxxxxxx> writes:
>>>
>>> [...]
>>>
>>>> Add WITH_DIRECT_CALLS support [3] (patch 3, 4)
>>>> ==============================================
>>>
>>> We've had some offlist discussions, so here's some input for a wider
>>> audience! Most importantly, this is for Palmer, so that this series is
>>> not merged until a proper BPF trampoline fix is in place.
>>>
>>> Note that what's currently usable from BPF trampoline *works*. It's
>>> when this series is added that it breaks.
>>>
>>> TL;DR This series adds DYNAMIC_FTRACE_WITH_DIRECT_CALLS, which enables
>>> fentry/fexit BPF trampoline support. Unfortunately the
>>> fexit/BPF_TRAMP_F_SKIP_FRAME parts of the RV BPF trampoline breaks
>>> with this addition, and need to be addressed *prior* merging this
>>> series. An easy way to reproduce, is just calling any of the kselftest
>>> tests that uses fexit patching.
>>>
>>> The issue is around the nop seld, and how a call is done; The nop sled
>>> (patchable-function-entry) size changed from 16B to 8B in commit
>>> 6724a76cff85 ("riscv: ftrace: Reduce the detour code size to half"), but
>>> BPF code still uses the old 16B. So it'll work for BPF programs, but not
>>> for regular kernel functions.
>>>
>>> An example:
>>>
>>> | ffffffff80fa4150 <bpf_fentry_test1>:
>>> | ffffffff80fa4150: 0001 nop
>>> | ffffffff80fa4152: 0001 nop
>>> | ffffffff80fa4154: 0001 nop
>>> | ffffffff80fa4156: 0001 nop
>>> | ffffffff80fa4158: 1141 add sp,sp,-16
>>> | ffffffff80fa415a: e422 sd s0,8(sp)
>>> | ffffffff80fa415c: 0800 add s0,sp,16
>>> | ffffffff80fa415e: 6422 ld s0,8(sp)
>>> | ffffffff80fa4160: 2505 addw a0,a0,1
>>> | ffffffff80fa4162: 0141 add sp,sp,16
>>> | ffffffff80fa4164: 8082 ret
>>>
>>> is patched to:
>>>
>>> | ffffffff80fa4150: f70c0297 auipc t0,-150208512
>>> | ffffffff80fa4154: eb0282e7 jalr t0,t0,-336
>>>
>>> The return address to bpf_fentry_test1 is stored in t0 at BPF
>>> trampoline entry. Return to the *parent* is in ra. The trampline has
>>> to deal with this.
>>>
>>> For BPF_TRAMP_F_SKIP_FRAME/CALL_ORIG, the BPF trampoline will skip too
>>> many bytes, and not correctly handle parent calls.
>>>
>>> Further; The BPF trampoline currently has a different way of patching
>>> the nops for BPF programs, than what ftrace does. That should be changed
>>> to match what ftrace does (auipc/jalr t0).
>>>
>>> To summarize:
>>> * Align BPF nop sled with patchable-function-entry: 8B.
>>> * Adapt BPF trampoline for 8B nop sleds.
>>> * Adapt BPF trampoline t0 return, ra parent scheme.
>>
>> Thanks for digging into this one, I agree we need to sort out the BPF
>> breakages before we merge this. Sounds like there's a rabbit hole here,
>> but hopefully we can get it sorted out.
>>
>> I've dropped this from patchwork and such, as we'll need at least
>> another spin.
>
> Palmer,
>
> The needed BPF patch is upstream in the bpf-next tree, and has been for
> a couple of weeks.
>
> I think this series is a candidate for RISC-V -next! It would help
> RISC-V BPF a lot in terms of completeness.
Palmer,
The needed fix for BPF is now in Linus' tree, commit 25ad10658dc1
("riscv, bpf: Adapt bpf trampoline to optimized riscv ftrace
framework"). IOW, this ftrace series can be merged now.
Björn
