> On Nov 6, 2023, at 1:07 PM, Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Fri, Nov 3, 2023 at 5:13 PM Song Liu <song@xxxxxxxxxx> wrote:
>>
>> Different types of bpf dynptr have different internal data storage.
>> Specifically, SKB and XDP type of dynptr may have non-continuous data.
>> Therefore, it is not always safe to directly access dynptr->data.
>>
>> Add __bpf_dynptr_data and __bpf_dynptr_data_rw to replace direct access to
>> dynptr->data.
>>
>> Update bpf_verify_pkcs7_signature to use __bpf_dynptr_data instead of
>> dynptr->data.
>>
>> Signed-off-by: Song Liu <song@xxxxxxxxxx>
>> ---
>> include/linux/bpf.h | 2 ++
>> kernel/bpf/helpers.c | 47 ++++++++++++++++++++++++++++++++++++++++
>> kernel/trace/bpf_trace.c | 12 ++++++----
>> 3 files changed, 57 insertions(+), 4 deletions(-)
>>
>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>> index b4825d3cdb29..eb84caf133df 100644
>> --- a/include/linux/bpf.h
>> +++ b/include/linux/bpf.h
>> @@ -1222,6 +1222,8 @@ enum bpf_dynptr_type {
>>
>> int bpf_dynptr_check_size(u32 size);
>> u32 __bpf_dynptr_size(const struct bpf_dynptr_kern *ptr);
>> +const void *__bpf_dynptr_data(const struct bpf_dynptr_kern *ptr, u32 len);
>> +void *__bpf_dynptr_data_rw(const struct bpf_dynptr_kern *ptr, u32 len);
>>
>> #ifdef CONFIG_BPF_JIT
>> int bpf_trampoline_link_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr);
>> diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c
>> index e46ac288a108..c569c4c43bde 100644
>> --- a/kernel/bpf/helpers.c
>> +++ b/kernel/bpf/helpers.c
>> @@ -2611,3 +2611,50 @@ static int __init kfunc_init(void)
>> }
>>
>> late_initcall(kfunc_init);
>> +
>> +/* Get a pointer to dynptr data up to len bytes for read only access. If
>> + * the dynptr doesn't have continuous data up to len bytes, return NULL.
>> + */
>> +const void *__bpf_dynptr_data(const struct bpf_dynptr_kern *ptr, u32 len)
>> +{
>> + enum bpf_dynptr_type type;
>> + int err;
>> +
>> + if (!ptr->data)
>> + return NULL;
>> +
>> + err = bpf_dynptr_check_off_len(ptr, 0, len);
>> + if (err)
>> + return NULL;
>> + type = bpf_dynptr_get_type(ptr);
>> +
>> + switch (type) {
>> + case BPF_DYNPTR_TYPE_LOCAL:
>> + case BPF_DYNPTR_TYPE_RINGBUF:
>> + return ptr->data + ptr->offset;
>> + case BPF_DYNPTR_TYPE_SKB:
>> + return skb_pointer_if_linear(ptr->data, ptr->offset, len);
>> + case BPF_DYNPTR_TYPE_XDP:
>> + {
>> + void *xdp_ptr = bpf_xdp_pointer(ptr->data, ptr->offset, len);
>> +
>> + if (IS_ERR_OR_NULL(xdp_ptr))
>> + return NULL;
>> + return xdp_ptr;
>> + }
>> + default:
>> + WARN_ONCE(true, "unknown dynptr type %d\n", type);
>> + return NULL;
>> + }
>> +}
>> +
>
> Song, you basically reimplemented bpf_dynptr_slice() but didn't unify
> the code. Now we have two almost identical non-trivial functions we'd
> need to update every time someone adds a new type of dynptr. Why not
> have common helper that does everything both bpf_dynptr_slice() kfunc
> needs and __bpf_dynptr_data() needs. And then call into it from both,
> keeping all the LOCAL vs RINGBUF vs SKB vs XDP logic in one place?
>
> Is there some problem unifying them?
Initially, I was thinking "buffer__opt == NULL && buffer__szk != 0" was
a problem for bpf_dynptr_slice(). And the buffer__opt == NULL case may
make a common helper more complicated. So I decided to not unify the two.
After a second look at it, I agree it shouldn't be a problem. And actually
we can do: (though you may argue against)
const void *__bpf_dynptr_data(const struct bpf_dynptr_kern *ptr, u32 len)
{
return bpf_dynptr_slice(ptr, 0, NULL, len);
}
As we are on this, shall we update bpf_dynptr_slice() to return
"const void *"? This is a little weird for buffer_opt != NULL, case as
buffer_opt is not const. But the compiler (clang) doesn't seem to
complain about it.
If we cannot have bpf_dynptr_slice() return const pointer, we will need
a little more casting for __bpf_dynptr_data().
Thanks,
Song
>
>> +/* Get a pointer to dynptr data up to len bytes for read write access. If
>> + * the dynptr doesn't have continuous data up to len bytes, or the dynptr
>> + * is read only, return NULL.
>> + */
