Re: [PATCH v3 bpf-next 2/7] bpf: add bpf_trace_vprintk helper

[Dave Marchevsky <davemarchevsky@fb.com>]

On 9/3/21 4:00 AM, Daniel Borkmann wrote:   
> On 8/28/21 7:20 AM, Dave Marchevsky wrote:
>> This helper is meant to be "bpf_trace_printk, but with proper vararg
>> support". Follow bpf_snprintf's example and take a u64 pseudo-vararg
>> array. Write to /sys/kernel/debug/tracing/trace_pipe using the same
>> mechanism as bpf_trace_printk.
>>
>> Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
> 
> lgtm, minor comments below:
> 
>> ---
>>   include/linux/bpf.h            |  1 +
>>   include/uapi/linux/bpf.h       |  9 ++++++
>>   kernel/bpf/core.c              |  5 ++++
>>   kernel/bpf/helpers.c           |  2 ++
>>   kernel/trace/bpf_trace.c       | 52 +++++++++++++++++++++++++++++++++-
>>   tools/include/uapi/linux/bpf.h |  9 ++++++
>>   6 files changed, 77 insertions(+), 1 deletion(-)
>>
>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>> index be8d57e6e78a..b6c45a6cbbba 100644
>> --- a/include/linux/bpf.h
>> +++ b/include/linux/bpf.h
>> @@ -1088,6 +1088,7 @@ bool bpf_prog_array_compatible(struct bpf_array *array, const struct bpf_prog *f
>>   int bpf_prog_calc_tag(struct bpf_prog *fp);
>>     const struct bpf_func_proto *bpf_get_trace_printk_proto(void);
>> +const struct bpf_func_proto *bpf_get_trace_vprintk_proto(void);
>>     typedef unsigned long (*bpf_ctx_copy_t)(void *dst, const void *src,
>>                       unsigned long off, unsigned long len);
>> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
>> index 791f31dd0abe..f171d4d33136 100644
>> --- a/include/uapi/linux/bpf.h
>> +++ b/include/uapi/linux/bpf.h
>> @@ -4877,6 +4877,14 @@ union bpf_attr {
>>    *        Get the struct pt_regs associated with **task**.
>>    *    Return
>>    *        A pointer to struct pt_regs.
>> + *
>> + * u64 bpf_trace_vprintk(const char *fmt, u32 fmt_size, const void *data, u32 data_len)
> 
> s/u64/long/
> 
>> + *    Description
>> + *        Behaves like **bpf_trace_printk**\ () helper, but takes an array of u64
> 
> nit: maybe for users it's more clear from description if you instead mention that data_len
> needs to be multiple of 8 bytes? Or somehow mention the relation with data more clearly
> resp. which shortcoming it addresses compared to bpf_trace_printk(), so developers can more
> easily parse it.

In a previous review pass, Andrii preferred having bpf_trace_vprintk's reference other helpers
instead of copy/pasting. So in v5 (patch 9) of this patchset I've added "multiple of 8 bytes"
to helper comments for bpf_seq_printf and bpf_snprintf. Added a sentence mentioning benefits
of vprintk over printk in v5 (patch 3).

>> + *        to format. Arguments are to be used as in **bpf_seq_printf**\ () helper.
>> + *    Return
>> + *        The number of bytes written to the buffer, or a negative error
>> + *        in case of failure.
>>    */
> [...]
>>       default:
>>           return NULL;
>>       }
>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>> index 10672ebc63b7..ea8358b0c748 100644
>> --- a/kernel/trace/bpf_trace.c
>> +++ b/kernel/trace/bpf_trace.c
>> @@ -398,7 +398,7 @@ static const struct bpf_func_proto bpf_trace_printk_proto = {
>>       .arg2_type    = ARG_CONST_SIZE,
>>   };
>>   -const struct bpf_func_proto *bpf_get_trace_printk_proto(void)
>> +static void __set_printk_clr_event(void)
>>   {
>>       /*
>>        * This program might be calling bpf_trace_printk,
>> @@ -410,10 +410,58 @@ const struct bpf_func_proto *bpf_get_trace_printk_proto(void)
>>        */
>>       if (trace_set_clr_event("bpf_trace", "bpf_trace_printk", 1))
>>           pr_warn_ratelimited("could not enable bpf_trace_printk events");
>> +}
>>   +const struct bpf_func_proto *bpf_get_trace_printk_proto(void)
>> +{
>> +    __set_printk_clr_event();
>>       return &bpf_trace_printk_proto;
>>   }
>>   +BPF_CALL_4(bpf_trace_vprintk, char *, fmt, u32, fmt_size, const void *, data,
>> +       u32, data_len)
>> +{
>> +    static char buf[BPF_TRACE_PRINTK_SIZE];
>> +    unsigned long flags;
>> +    int ret, num_args;
>> +    u32 *bin_args;
>> +
>> +    if (data_len & 7 || data_len > MAX_BPRINTF_VARARGS * 8 ||
>> +        (data_len && !data))
>> +        return -EINVAL;
>> +    num_args = data_len / 8;
>> +
>> +    ret = bpf_bprintf_prepare(fmt, fmt_size, data, &bin_args, num_args);
>> +    if (ret < 0)
>> +        return ret;
> 
> Given you have ARG_PTR_TO_MEM_OR_NULL for data, does this gracefully handle the
> case where you pass in fmt string containing e.g. %ps but data being NULL? From
> reading bpf_bprintf_prepare() looks like it does just fine, but might be nice
> to explicitly add a tiny selftest case for it while you're at it.
> 
>> +    raw_spin_lock_irqsave(&trace_printk_lock, flags);
>> +    ret = bstr_printf(buf, sizeof(buf), fmt, bin_args);
>> +
>> +    trace_bpf_trace_printk(buf);
>> +    raw_spin_unlock_irqrestore(&trace_printk_lock, flags);
>> +
>> +    bpf_bprintf_cleanup();
>> +
>> +    return ret;
>> +}
> 
> Thanks,
> Daniel