classic BPF has a restriction that last insn is always BPF_RET.
eBPF doesn't have BPF_RET instruction and this restriction.
It has BPF_EXIT insn which can appear anywhere in the program
one or more times and it doesn't have to be last insn.
Fix eBPF JIT to emit epilogue when first BPF_EXIT is seen
and all other BPF_EXIT instructions will be emitted as jump.
Signed-off-by: Alexei Starovoitov <[email protected]>
---
Note, this bug is applicable only to native eBPF programs
which first were introduced in 3.18, so no need to send it
to stable and therefore no 'Fixes' tag.
arm64 JIT has the same problem, but the fix is not as trivial,
so will be done as separate patch.
Since 3.18 can only load eBPF programs and cannot execute them,
this patch can even be done in net-next only, but I think it's worth
to apply it to 3.18(net), so that JITed output for native eBPF
programs is correct when bpf syscall loads it with net.core.bpf_jit_enable=2
arch/x86/net/bpf_jit_comp.c | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
index 3f62734..7e90244 100644
--- a/arch/x86/net/bpf_jit_comp.c
+++ b/arch/x86/net/bpf_jit_comp.c
@@ -178,7 +178,7 @@ static void jit_fill_hole(void *area, unsigned int size)
}
struct jit_context {
- unsigned int cleanup_addr; /* epilogue code offset */
+ int cleanup_addr; /* epilogue code offset */
bool seen_ld_abs;
};
@@ -192,6 +192,7 @@ static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
struct bpf_insn *insn = bpf_prog->insnsi;
int insn_cnt = bpf_prog->len;
bool seen_ld_abs = ctx->seen_ld_abs | (oldproglen == 0);
+ bool seen_exit = false;
u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
int i;
int proglen = 0;
@@ -854,10 +855,11 @@ common_load:
goto common_load;
case BPF_JMP | BPF_EXIT:
- if (i != insn_cnt - 1) {
+ if (seen_exit) {
jmp_offset = ctx->cleanup_addr - addrs[i];
goto emit_jmp;
}
+ seen_exit = true;
/* update cleanup_addr */
ctx->cleanup_addr = proglen;
/* mov rbx, qword ptr [rbp-X] */
--
1.7.9.5
On 11/27/2014 06:02 AM, Alexei Starovoitov wrote:
> classic BPF has a restriction that last insn is always BPF_RET.
> eBPF doesn't have BPF_RET instruction and this restriction.
> It has BPF_EXIT insn which can appear anywhere in the program
> one or more times and it doesn't have to be last insn.
> Fix eBPF JIT to emit epilogue when first BPF_EXIT is seen
> and all other BPF_EXIT instructions will be emitted as jump.
>
> Signed-off-by: Alexei Starovoitov <[email protected]>
> ---
> Note, this bug is applicable only to native eBPF programs
> which first were introduced in 3.18, so no need to send it
> to stable and therefore no 'Fixes' tag.
Btw, even if it's not sent to -stable, a 'Fixes:' tag is useful
information for backporting and regression tracking, preferably
always mentioned where it can clearly be identified.
> arm64 JIT has the same problem, but the fix is not as trivial,
> so will be done as separate patch.
>
> Since 3.18 can only load eBPF programs and cannot execute them,
> this patch can even be done in net-next only, but I think it's worth
> to apply it to 3.18(net), so that JITed output for native eBPF
> programs is correct when bpf syscall loads it with net.core.bpf_jit_enable=2
Yes, sounds good to me, the condition insn_cnt - 1 is still held
with BPF to eBPF transformations.
> arch/x86/net/bpf_jit_comp.c | 6 ++++--
> 1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
> index 3f62734..7e90244 100644
> --- a/arch/x86/net/bpf_jit_comp.c
> +++ b/arch/x86/net/bpf_jit_comp.c
> @@ -178,7 +178,7 @@ static void jit_fill_hole(void *area, unsigned int size)
> }
>
> struct jit_context {
> - unsigned int cleanup_addr; /* epilogue code offset */
> + int cleanup_addr; /* epilogue code offset */
Why this type change here? This seems a bit out of context (I would
have expected a mention of this in the commit message, otherwise).
> bool seen_ld_abs;
> };
>
> @@ -192,6 +192,7 @@ static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
> struct bpf_insn *insn = bpf_prog->insnsi;
> int insn_cnt = bpf_prog->len;
> bool seen_ld_abs = ctx->seen_ld_abs | (oldproglen == 0);
> + bool seen_exit = false;
> u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
> int i;
> int proglen = 0;
> @@ -854,10 +855,11 @@ common_load:
> goto common_load;
>
> case BPF_JMP | BPF_EXIT:
> - if (i != insn_cnt - 1) {
> + if (seen_exit) {
> jmp_offset = ctx->cleanup_addr - addrs[i];
> goto emit_jmp;
> }
> + seen_exit = true;
> /* update cleanup_addr */
> ctx->cleanup_addr = proglen;
> /* mov rbx, qword ptr [rbp-X] */
>
On Thu, Nov 27, 2014 at 1:52 AM, Daniel Borkmann <[email protected]> wrote:
> On 11/27/2014 06:02 AM, Alexei Starovoitov wrote:
>>
>> classic BPF has a restriction that last insn is always BPF_RET.
>> eBPF doesn't have BPF_RET instruction and this restriction.
>> It has BPF_EXIT insn which can appear anywhere in the program
>> one or more times and it doesn't have to be last insn.
>> Fix eBPF JIT to emit epilogue when first BPF_EXIT is seen
>> and all other BPF_EXIT instructions will be emitted as jump.
>>
>> Signed-off-by: Alexei Starovoitov <[email protected]>
>> ---
>> Note, this bug is applicable only to native eBPF programs
>> which first were introduced in 3.18, so no need to send it
>> to stable and therefore no 'Fixes' tag.
>
>
> Btw, even if it's not sent to -stable, a 'Fixes:' tag is useful
> information for backporting and regression tracking, preferably
> always mentioned where it can clearly be identified.
Well I didn't mention it, as I said, because I don't think it
needs backporting. Otherwise with the tag the tools might
pick it up automatically? Just a guess.
Anyway:
Fixes: 622582786c9e ("net: filter: x86: internal BPF JIT")
>> arm64 JIT has the same problem, but the fix is not as trivial,
>> so will be done as separate patch.
>>
>> Since 3.18 can only load eBPF programs and cannot execute them,
>> this patch can even be done in net-next only, but I think it's worth
>> to apply it to 3.18(net), so that JITed output for native eBPF
>> programs is correct when bpf syscall loads it with
>> net.core.bpf_jit_enable=2
>
>
> Yes, sounds good to me, the condition insn_cnt - 1 is still held
> with BPF to eBPF transformations.
Correct. That's what I meant that prior to 3.18 it's not needed.
and 'insn_cnt - 1' condition will still hold for classic in the future.
>> arch/x86/net/bpf_jit_comp.c | 6 ++++--
>> 1 file changed, 4 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
>> index 3f62734..7e90244 100644
>> --- a/arch/x86/net/bpf_jit_comp.c
>> +++ b/arch/x86/net/bpf_jit_comp.c
>> @@ -178,7 +178,7 @@ static void jit_fill_hole(void *area, unsigned int
>> size)
>> }
>>
>> struct jit_context {
>> - unsigned int cleanup_addr; /* epilogue code offset */
>> + int cleanup_addr; /* epilogue code offset */
>
>
> Why this type change here? This seems a bit out of context (I would
> have expected a mention of this in the commit message, otherwise).
Ok. Will respin with updated commit msg.
The reason for signed is the following:
jmp offset to epilogue is computed as:
jmp_offset = ctx->cleanup_addr - addrs[i]
when cleanup_addr was always last insn it wasn't a problem,
since result of subtraction was positive.
Now, since epilogue will be in the middle of JITed
code the jmps to epilogue may be negative, so
signed int is need to do the math correctly.
In other words, it should be:
(long long) ((int)20 - (int)30)
instead of:
(long long) ((unsigned int)20 - (int)30)
On 11/28/2014 06:55 AM, Alexei Starovoitov wrote:
> On Thu, Nov 27, 2014 at 1:52 AM, Daniel Borkmann <[email protected]> wrote:
>> On 11/27/2014 06:02 AM, Alexei Starovoitov wrote:
>>>
>>> classic BPF has a restriction that last insn is always BPF_RET.
>>> eBPF doesn't have BPF_RET instruction and this restriction.
>>> It has BPF_EXIT insn which can appear anywhere in the program
>>> one or more times and it doesn't have to be last insn.
>>> Fix eBPF JIT to emit epilogue when first BPF_EXIT is seen
>>> and all other BPF_EXIT instructions will be emitted as jump.
>>>
>>> Signed-off-by: Alexei Starovoitov <[email protected]>
>>> ---
>>> Note, this bug is applicable only to native eBPF programs
>>> which first were introduced in 3.18, so no need to send it
>>> to stable and therefore no 'Fixes' tag.
>>
>> Btw, even if it's not sent to -stable, a 'Fixes:' tag is useful
>> information for backporting and regression tracking, preferably
>> always mentioned where it can clearly be identified.
>
> Well I didn't mention it, as I said, because I don't think it
> needs backporting. Otherwise with the tag the tools might
> pick it up automatically? Just a guess.
No, Dave selects -stable material on a case-by-case basis and bundles
it up eventually; after -net was merged, it's then pushed to -stable
by himself (see Documentation/networking/netdev-FAQ.txt +114). So the
comment below "---" is absolutely okay.
It can well be, that some people/companies cannot switch for various
reasons immediately to the next kernels, but nevertheless would like
to have a certain features included, so generally regression tracking
via 'Fixes:' tag is extremely useful/valuable. ;)
> Fixes: 622582786c9e ("net: filter: x86: internal BPF JIT")
...
>> Why this type change here? This seems a bit out of context (I would
>> have expected a mention of this in the commit message, otherwise).
>
> The reason for signed is the following:
> jmp offset to epilogue is computed as:
> jmp_offset = ctx->cleanup_addr - addrs[i]
> when cleanup_addr was always last insn it wasn't a problem,
> since result of subtraction was positive.
> Now, since epilogue will be in the middle of JITed
> code the jmps to epilogue may be negative
Ok, thanks for the clarification, Alexei.