| Message ID | 20221103092118.248600-3-yangjihong1@huawei.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | bpf: Support kernel function call in 32-bit ARM | expand |
On Thu, Nov 03, 2022 at 05:21:16PM +0800, Yang Jihong wrote: > The error code -EACCES is returned when bpf prog is tested in 32-bit environment, > This is because bpf_object__relocate modifies the instruction to change memory > size to 4 bytes, as shown in the following messages: > > libbpf: prog 'kfunc_call_test1': relo #2: matching candidate #0 <byte_off> [18342] struct __sk_buff.sk (0:30:0 @ offset 168) > libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) off 168 -> 168 > libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) mem_sz 8 -> 4 > > As a result, the bpf_skb_is_valid_access check fails. For 32-bit architecture, > unnecessary checks need to be deleted. Isn't the purpose of this check to ensure that the entire pointer is written, and BPF can't write half of it? > case offsetof(struct __sk_buff, sk): > - if (type == BPF_WRITE || size != sizeof(__u64)) > - return false; Wouldn't "(size != sizeof(struct bpf_sock *) && size != sizeof(__u64))" be more appropriate here, so 32-bit can only write the 32-bit pointer or the full 64-bit value, and 64-bit can only write the 64-bit pointer? Or is there a reason not to? bpf folk?
On Thu, Nov 3, 2022 at 4:23 AM Russell King (Oracle) <linux@armlinux.org.uk> wrote: > > On Thu, Nov 03, 2022 at 05:21:16PM +0800, Yang Jihong wrote: > > The error code -EACCES is returned when bpf prog is tested in 32-bit environment, > > This is because bpf_object__relocate modifies the instruction to change memory > > size to 4 bytes, as shown in the following messages: > > > > libbpf: prog 'kfunc_call_test1': relo #2: matching candidate #0 <byte_off> [18342] struct __sk_buff.sk (0:30:0 @ offset 168) > > libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) off 168 -> 168 > > libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) mem_sz 8 -> 4 > > > > As a result, the bpf_skb_is_valid_access check fails. For 32-bit architecture, > > unnecessary checks need to be deleted. > > Isn't the purpose of this check to ensure that the entire pointer is > written, and BPF can't write half of it? > > > > case offsetof(struct __sk_buff, sk): > > - if (type == BPF_WRITE || size != sizeof(__u64)) > > - return false; > > Wouldn't "(size != sizeof(struct bpf_sock *) && size != sizeof(__u64))" > be more appropriate here, so 32-bit can only write the 32-bit pointer > or the full 64-bit value, and 64-bit can only write the 64-bit pointer? > Or is there a reason not to? bpf folk? You're correct. The patch is completely wrong. The bug is elsewhere.
On Thu, Nov 3, 2022 at 11:15 AM Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote: > > On Thu, Nov 3, 2022 at 4:23 AM Russell King (Oracle) > <linux@armlinux.org.uk> wrote: > > > > On Thu, Nov 03, 2022 at 05:21:16PM +0800, Yang Jihong wrote: > > > The error code -EACCES is returned when bpf prog is tested in 32-bit environment, > > > This is because bpf_object__relocate modifies the instruction to change memory > > > size to 4 bytes, as shown in the following messages: > > > > > > libbpf: prog 'kfunc_call_test1': relo #2: matching candidate #0 <byte_off> [18342] struct __sk_buff.sk (0:30:0 @ offset 168) > > > libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) off 168 -> 168 > > > libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) mem_sz 8 -> 4 > > > > > > As a result, the bpf_skb_is_valid_access check fails. For 32-bit architecture, > > > unnecessary checks need to be deleted. > > > > Isn't the purpose of this check to ensure that the entire pointer is > > written, and BPF can't write half of it? > > > > > > > case offsetof(struct __sk_buff, sk): > > > - if (type == BPF_WRITE || size != sizeof(__u64)) > > > - return false; > > > > Wouldn't "(size != sizeof(struct bpf_sock *) && size != sizeof(__u64))" > > be more appropriate here, so 32-bit can only write the 32-bit pointer > > or the full 64-bit value, and 64-bit can only write the 64-bit pointer? > > Or is there a reason not to? bpf folk? > > You're correct. The patch is completely wrong. > The bug is elsewhere. So I looked at this a bit (and replied to the old version of this patch). What happens in the kernel is that we expect 64-bit load but rewrite it to 32-bit load on 32-bit architectures (because we just use sizeof(struct sk_buff, sk) which is 4 bytes on 32-bit arch. The problem here is that libbpf adjusts such pointer accesses from 8-byte read to 4-byte reads for preserve_access_index (because libbpf sees that pointer is really 4 byte long), which is what we actually want in the general case. Here the assumption was made before CO-RE that __sk_buff is a stable (and fake) UAPI and the correct BPF program will access sk as a 64-bit pointer because BPF-side pointers always appear as 64-bit. But from a correctness standpoint I think it should be fine to enable both 32- and 64-bit loads for such pointers in __sk_buff for 32-bit host arch. This will work well with CO-RE and will be correctly rewritten to 32-bit or 64-bit accesses, depending on host architecture. We should still reject 32-bit load on 64-bit host arch, though.
On Fri, Nov 4, 2022 at 3:43 PM Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote: > > On Thu, Nov 3, 2022 at 11:15 AM Alexei Starovoitov > <alexei.starovoitov@gmail.com> wrote: > > > > On Thu, Nov 3, 2022 at 4:23 AM Russell King (Oracle) > > <linux@armlinux.org.uk> wrote: > > > > > > On Thu, Nov 03, 2022 at 05:21:16PM +0800, Yang Jihong wrote: > > > > The error code -EACCES is returned when bpf prog is tested in 32-bit environment, > > > > This is because bpf_object__relocate modifies the instruction to change memory > > > > size to 4 bytes, as shown in the following messages: > > > > > > > > libbpf: prog 'kfunc_call_test1': relo #2: matching candidate #0 <byte_off> [18342] struct __sk_buff.sk (0:30:0 @ offset 168) > > > > libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) off 168 -> 168 > > > > libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) mem_sz 8 -> 4 > > > > > > > > As a result, the bpf_skb_is_valid_access check fails. For 32-bit architecture, > > > > unnecessary checks need to be deleted. > > > > > > Isn't the purpose of this check to ensure that the entire pointer is > > > written, and BPF can't write half of it? > > > > > > > > > > case offsetof(struct __sk_buff, sk): > > > > - if (type == BPF_WRITE || size != sizeof(__u64)) > > > > - return false; > > > > > > Wouldn't "(size != sizeof(struct bpf_sock *) && size != sizeof(__u64))" > > > be more appropriate here, so 32-bit can only write the 32-bit pointer > > > or the full 64-bit value, and 64-bit can only write the 64-bit pointer? > > > Or is there a reason not to? bpf folk? > > > > You're correct. The patch is completely wrong. > > The bug is elsewhere. > > So I looked at this a bit (and replied to the old version of this > patch). What happens in the kernel is that we expect 64-bit load but > rewrite it to 32-bit load on 32-bit architectures (because we just use > sizeof(struct sk_buff, sk) which is 4 bytes on 32-bit arch. > > The problem here is that libbpf adjusts such pointer accesses from > 8-byte read to 4-byte reads for preserve_access_index (because libbpf > sees that pointer is really 4 byte long), which is what we actually > want in the general case. Here the assumption was made before CO-RE > that __sk_buff is a stable (and fake) UAPI and the correct BPF program > will access sk as a 64-bit pointer because BPF-side pointers always > appear as 64-bit. > > But from a correctness standpoint I think it should be fine to enable > both 32- and 64-bit loads for such pointers in __sk_buff for 32-bit > host arch. This will work well with CO-RE and will be correctly > rewritten to 32-bit or 64-bit accesses, depending on host > architecture. > > We should still reject 32-bit load on 64-bit host arch, though. Replied in the other thread as well :) The CO_RE screws up access here. Since it's a load of a pointer the verifier has to see it as a 8-byte load. When CO-RE converts it to 4 byte the verifier won't recognize it as a pointer load anymore. We cannot easily convert 64-bit BPF ISA into 32-bit. It's a massive amount of work.
Hello, On 2022/11/3 19:23, Russell King (Oracle) wrote: > On Thu, Nov 03, 2022 at 05:21:16PM +0800, Yang Jihong wrote: >> The error code -EACCES is returned when bpf prog is tested in 32-bit environment, >> This is because bpf_object__relocate modifies the instruction to change memory >> size to 4 bytes, as shown in the following messages: >> >> libbpf: prog 'kfunc_call_test1': relo #2: matching candidate #0 <byte_off> [18342] struct __sk_buff.sk (0:30:0 @ offset 168) >> libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) off 168 -> 168 >> libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) mem_sz 8 -> 4 >> >> As a result, the bpf_skb_is_valid_access check fails. For 32-bit architecture, >> unnecessary checks need to be deleted. > > Isn't the purpose of this check to ensure that the entire pointer is > written, and BPF can't write half of it? > > >> case offsetof(struct __sk_buff, sk): >> - if (type == BPF_WRITE || size != sizeof(__u64)) >> - return false; > > Wouldn't "(size != sizeof(struct bpf_sock *) && size != sizeof(__u64))" > be more appropriate here, so 32-bit can only write the 32-bit pointer > or the full 64-bit value, and 64-bit can only write the 64-bit pointer? > Or is there a reason not to? bpf folk? > Thanks for the detailed proposals, will fix it in next version. Thanks, Yang
Hello, On 2022/11/5 7:37, Alexei Starovoitov wrote: > On Fri, Nov 4, 2022 at 3:43 PM Andrii Nakryiko > <andrii.nakryiko@gmail.com> wrote: >> >> On Thu, Nov 3, 2022 at 11:15 AM Alexei Starovoitov >> <alexei.starovoitov@gmail.com> wrote: >>> >>> On Thu, Nov 3, 2022 at 4:23 AM Russell King (Oracle) >>> <linux@armlinux.org.uk> wrote: >>>> >>>> On Thu, Nov 03, 2022 at 05:21:16PM +0800, Yang Jihong wrote: >>>>> The error code -EACCES is returned when bpf prog is tested in 32-bit environment, >>>>> This is because bpf_object__relocate modifies the instruction to change memory >>>>> size to 4 bytes, as shown in the following messages: >>>>> >>>>> libbpf: prog 'kfunc_call_test1': relo #2: matching candidate #0 <byte_off> [18342] struct __sk_buff.sk (0:30:0 @ offset 168) >>>>> libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) off 168 -> 168 >>>>> libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) mem_sz 8 -> 4 >>>>> >>>>> As a result, the bpf_skb_is_valid_access check fails. For 32-bit architecture, >>>>> unnecessary checks need to be deleted. >>>> >>>> Isn't the purpose of this check to ensure that the entire pointer is >>>> written, and BPF can't write half of it? >>>> >>>> >>>>> case offsetof(struct __sk_buff, sk): >>>>> - if (type == BPF_WRITE || size != sizeof(__u64)) >>>>> - return false; >>>> >>>> Wouldn't "(size != sizeof(struct bpf_sock *) && size != sizeof(__u64))" >>>> be more appropriate here, so 32-bit can only write the 32-bit pointer >>>> or the full 64-bit value, and 64-bit can only write the 64-bit pointer? >>>> Or is there a reason not to? bpf folk? >>> >>> You're correct. The patch is completely wrong. >>> The bug is elsewhere. >> >> So I looked at this a bit (and replied to the old version of this >> patch). What happens in the kernel is that we expect 64-bit load but >> rewrite it to 32-bit load on 32-bit architectures (because we just use >> sizeof(struct sk_buff, sk) which is 4 bytes on 32-bit arch. >> >> The problem here is that libbpf adjusts such pointer accesses from >> 8-byte read to 4-byte reads for preserve_access_index (because libbpf >> sees that pointer is really 4 byte long), which is what we actually >> want in the general case. Here the assumption was made before CO-RE >> that __sk_buff is a stable (and fake) UAPI and the correct BPF program >> will access sk as a 64-bit pointer because BPF-side pointers always >> appear as 64-bit. >> >> But from a correctness standpoint I think it should be fine to enable >> both 32- and 64-bit loads for such pointers in __sk_buff for 32-bit >> host arch. This will work well with CO-RE and will be correctly >> rewritten to 32-bit or 64-bit accesses, depending on host >> architecture. >> >> We should still reject 32-bit load on 64-bit host arch, though. > > Replied in the other thread as well :) > The CO_RE screws up access here. > Since it's a load of a pointer the verifier has to see it as a 8-byte load. > When CO-RE converts it to 4 byte the verifier won't recognize it > as a pointer load anymore. > We cannot easily convert 64-bit BPF ISA into 32-bit. > It's a massive amount of work. > . From the above discussion, there are two different solutions: 1. Modify bpf_skb_is_valid_access to ensure that 32-bit can only load the 32-bit pointer or the full 64-bit value, and 64-bit can only load the 64-bit pointer 2. Modify libbpf, CO_RE skips adjust load's mem size and retains the 8-byte load. The two fixes will be added in the next version. Please review the solution to be selected. Thanks, Yang
diff --git a/net/core/filter.c b/net/core/filter.c index bb0136e7a8e4..eab7ce89740c 100644 --- a/net/core/filter.c +++ b/net/core/filter.c @@ -8269,8 +8269,6 @@ static bool bpf_skb_is_valid_access(int off, int size, enum bpf_access_type type return false; break; case offsetof(struct __sk_buff, sk): - if (type == BPF_WRITE || size != sizeof(__u64)) - return false; info->reg_type = PTR_TO_SOCK_COMMON_OR_NULL; break; case offsetof(struct __sk_buff, tstamp_type):
The error code -EACCES is returned when bpf prog is tested in 32-bit environment, This is because bpf_object__relocate modifies the instruction to change memory size to 4 bytes, as shown in the following messages: libbpf: prog 'kfunc_call_test1': relo #2: matching candidate #0 <byte_off> [18342] struct __sk_buff.sk (0:30:0 @ offset 168) libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) off 168 -> 168 libbpf: prog 'kfunc_call_test1': relo #2: patched insn #1 (LDX/ST/STX) mem_sz 8 -> 4 As a result, the bpf_skb_is_valid_access check fails. For 32-bit architecture, unnecessary checks need to be deleted. Signed-off-by: Yang Jihong <yangjihong1@huawei.com> --- net/core/filter.c | 2 -- 1 file changed, 2 deletions(-)