Re: [PATCH bpf-next 1/5] bpf: Add kfunc bpf_get_file_xattr

[Song Liu <songliubraving@meta.com>]

> On Oct 17, 2023, at 2:52 PM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
> 
> On Tue, Oct 17, 2023 at 1:31 PM Song Liu <songliubraving@meta.com> wrote:
>> 
>> 
>> 
>>> On Oct 17, 2023, at 11:58 AM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
>>> 
>>> On Fri, Oct 13, 2023 at 11:29 AM Song Liu <song@kernel.org> wrote:
>>>> 
>>>> This kfunc can be used to read xattr of a file.
>>>> 
>>>> Since vfs_getxattr() requires null-terminated string as input "name", a new
>>>> helper bpf_dynptr_is_string() is added to check the input before calling
>>>> vfs_getxattr().
>>>> 
>>>> Signed-off-by: Song Liu <song@kernel.org>
>>>> ---
>>>> include/linux/bpf.h      | 12 +++++++++++
>>>> kernel/trace/bpf_trace.c | 44 ++++++++++++++++++++++++++++++++++++++++
>>>> 2 files changed, 56 insertions(+)
>>>> 
>>>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>>>> index 61bde4520f5c..f14fae45e13d 100644
>>>> --- a/include/linux/bpf.h
>>>> +++ b/include/linux/bpf.h
>>>> @@ -2472,6 +2472,13 @@ static inline bool has_current_bpf_ctx(void)
>>>>       return !!current->bpf_ctx;
>>>> }
>>>> 
>>>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
>>> 
>>> is_zero_terminated would be more accurate? though there is nothing
>>> really dynptr-specific here...
>> 
>> is_zero_terminated sounds better.
>> 
>>> 
>>>> +{
>>>> +       char *str = ptr->data;
>>>> +
>>>> +       return str[__bpf_dynptr_size(ptr) - 1] == '\0';
>>>> +}
>>>> +
>>>> void notrace bpf_prog_inc_misses_counter(struct bpf_prog *prog);
>>>> 
>>>> void bpf_dynptr_init(struct bpf_dynptr_kern *ptr, void *data,
>>>> @@ -2708,6 +2715,11 @@ static inline bool has_current_bpf_ctx(void)
>>>>       return false;
>>>> }
>>>> 
>>>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
>>>> +{
>>>> +       return false;
>>>> +}
>>>> +
>>>> static inline void bpf_prog_inc_misses_counter(struct bpf_prog *prog)
>>>> {
>>>> }
>>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>>> index df697c74d519..946268574e05 100644
>>>> --- a/kernel/trace/bpf_trace.c
>>>> +++ b/kernel/trace/bpf_trace.c
>>>> @@ -24,6 +24,7 @@
>>>> #include <linux/key.h>
>>>> #include <linux/verification.h>
>>>> #include <linux/namei.h>
>>>> +#include <linux/fileattr.h>
>>>> 
>>>> #include <net/bpf_sk_storage.h>
>>>> 
>>>> @@ -1429,6 +1430,49 @@ static int __init bpf_key_sig_kfuncs_init(void)
>>>> late_initcall(bpf_key_sig_kfuncs_init);
>>>> #endif /* CONFIG_KEYS */
>>>> 
>>>> +/* filesystem kfuncs */
>>>> +__diag_push();
>>>> +__diag_ignore_all("-Wmissing-prototypes",
>>>> +                 "kfuncs which will be used in BPF programs");
>>>> +
>>>> +/**
>>>> + * bpf_get_file_xattr - get xattr of a file
>>>> + * @name_ptr: name of the xattr
>>>> + * @value_ptr: output buffer of the xattr value
>>>> + *
>>>> + * Get xattr *name_ptr* of *file* and store the output in *value_ptr*.
>>>> + *
>>>> + * Return: 0 on success, a negative value on error.
>>>> + */
>>>> +__bpf_kfunc int bpf_get_file_xattr(struct file *file, struct bpf_dynptr_kern *name_ptr,
>>>> +                                  struct bpf_dynptr_kern *value_ptr)
>>>> +{
>>>> +       if (!bpf_dynptr_is_string(name_ptr))
>>>> +               return -EINVAL;
>>> 
>>> so dynptr can be invalid and name_ptr->data will be NULL, you should
>>> account for that
>> 
>> We can add a NULL check (or size check) here.
> 
> there must be some helper to check if dynptr is valid, let's use that
> instead of NULL checks

Yeah, we can use bpf_dynptr_is_null(). 

> 
>> 
>>> 
>>> and there could also be special dynptrs that don't have contiguous
>>> memory region, so somehow you'd need to take care of that as well
>> 
>> We can require the dynptr to be BPF_DYNPTR_TYPE_LOCAL. I don't think
>> we need this for dynptr of skb or xdp. Would this be sufficient?
> 
> well, to keep thing simple we can have a simple internal helper API
> that will tell if it's safe to assume that dynptr memory is contiguous
> and it's ok to use dynptr memory. But still, you shouldn't access data
> pointer directly, there must be some helper for that. Please check. It
> has to take into account offset and stuff like that.

Yeah, we can use bpf_dynptr_write(), which is a helper (not kfunc). 

> 
> Also, and separately from that, we should think about providing a
> bpf_dynptr_slice()-like helper that will accept a fixed-sized
> temporary buffer and return pointer to either actual memory or copy
> non-contiguous memory into that buffer. That will make sure you can
> use any dynptr as a source of data, and only pay the price of memory
> copy in rare cases where it's necessary

I don't quite follow here. Currently, we have 

bpf_dynptr_data()
bpf_dynptr_slice()
bpf_dynptr_slice_rdwr()
bpf_dynptr_write()

AFAICT, they are sufficient to cover existing use cases (and the new 
use case we are adding in this set). What's the new kfunc are you
thinking about?

Thanks,
Song