On 11/7/23 06:23, Andrii Nakryiko wrote:
> On Fri, Jul 7, 2023 at 9:08 PM Leon Hwang <hffilwlqm@xxxxxxxxx> wrote:
>>
>> Currently, excluding verifier, users are unable to obtain detailed error
>> information when issues occur in BPF syscall.
>>
>> To overcome this limitation, bpf generic log is introduced to provide
>> error details similar to the verifier. This enhancement will enable the
>> reporting of error details along with the corresponding errno in BPF
>> syscall.
>>
>> Essentially, bpf generic log functions as a mechanism similar to netlink,
>> enabling the transmission of error messages to user space. This
>> mechanism greatly enhances the usability of BPF syscall by allowing
>> users to access comprehensive error messages instead of relying solely
>> on errno.
>>
>> This patch specifically addresses the error reporting in dev_xdp_attach()
>> . With this patch, the error messages will be transferred to user space
>> like the netlink approach. Hence, users will be able to check the error
>> message along with the errno.
>>
>> Signed-off-by: Leon Hwang <hffilwlqm@xxxxxxxxx>
>> ---
>> include/linux/bpf.h | 30 ++++++++++++++++++++++++++++++
>> include/uapi/linux/bpf.h | 6 ++++++
>> kernel/bpf/log.c | 33 +++++++++++++++++++++++++++++++++
>> net/core/dev.c | 11 ++++++++++-
>> tools/include/uapi/linux/bpf.h | 6 ++++++
>> 5 files changed, 85 insertions(+), 1 deletion(-)
>>
>
> Just curious, what's wrong with struct bpf_verifier_log for
> implementing "generic log"? bpf_log, bpf_vlog_reset, bpf_vlog_finalize
> are quite generic, I think. Why invent yet another structure? Existing
> code and struct support rotating log, if necessary.
>
Sorry for the misguiding patch. This patch should not be submitted in v2.
>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>> index 360433f14496a..7d2124a537943 100644
>> --- a/include/linux/bpf.h
>> +++ b/include/linux/bpf.h
>> @@ -3107,4 +3107,34 @@ static inline gfp_t bpf_memcg_flags(gfp_t flags)
>> return flags;
>> }
>>
>> +#define BPF_GENERIC_TMP_LOG_SIZE 256
>> +
>> +struct bpf_generic_log {
>> + char kbuf[BPF_GENERIC_TMP_LOG_SIZE];
>> + char __user *ubuf;
>> + u32 len_used;
>> + u32 len_total;
>> +};
>> +
>> +__printf(2, 3) void bpf_generic_log_write(struct bpf_generic_log *log,
>> + const char *fmt, ...);
>> +static inline void bpf_generic_log_init(struct bpf_generic_log *log,
>> + const struct bpf_generic_user_log *ulog)
>> +{
>> + log->ubuf = (char __user *) (unsigned long) ulog->log_buf;
>> + log->len_total = ulog->log_size;
>> + log->len_used = 0;
>> +}
>> +
>> +#define BPF_GENERIC_LOG_WRITE(log, ulog, fmt, ...) do { \
>> + const char *____fmt = (fmt); \
>> + bpf_generic_log_init(log, ulog); \
>> + bpf_generic_log_write(log, ____fmt, ##__VA_ARGS__); \
>> +} while (0)
>> +
>> +#define BPF_GENERIC_ULOG_WRITE(ulog, fmt, ...) do { \
>> + struct bpf_generic_log ____log; \
>> + BPF_GENERIC_LOG_WRITE(&____log, ulog, fmt, ##__VA_ARGS__); \
>> +} while (0)
>> +
>> #endif /* _LINUX_BPF_H */
>> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
>> index 60a9d59beeabb..34fa334938ba5 100644
>> --- a/include/uapi/linux/bpf.h
>> +++ b/include/uapi/linux/bpf.h
>> @@ -1318,6 +1318,11 @@ struct bpf_stack_build_id {
>> };
>> };
>>
>> +struct bpf_generic_user_log {
>> + __aligned_u64 log_buf; /* user supplied buffer */
>> + __u32 log_size; /* size of user buffer */
>> +};
>> +
>> #define BPF_OBJ_NAME_LEN 16U
>>
>> union bpf_attr {
>> @@ -1544,6 +1549,7 @@ union bpf_attr {
>> };
>> __u32 attach_type; /* attach type */
>> __u32 flags; /* extra flags */
>> + struct bpf_generic_user_log log; /* user log */
>
> I think explicit triplet of log_level (should be log_flags, but too
> late, probably), log_size, and log_buf is less error prone and more in
> sync with other two commands that accept log (BPF_PROG_LOAD and
> BPF_BTF_LOAD).
>
Better to keep same with BPF_PROG_LOAD and BPF_BTF_LOAD.
But considering the suggestion from Daniel Xu and Alexei Starovoitov,
it's better to add a tracepoint to expose the error message from
dev_xdp_attach(), without breaking uapi.
>> union {
>> __u32 target_btf_id; /* btf_id of target to attach to */
>> struct {
>> diff --git a/kernel/bpf/log.c b/kernel/bpf/log.c
>> index 850494423530e..be56b153bbf0b 100644
>> --- a/kernel/bpf/log.c
>> +++ b/kernel/bpf/log.c
>> @@ -325,3 +325,36 @@ __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
>> va_end(args);
>> }
>> EXPORT_SYMBOL_GPL(bpf_log);
>> +
>> +static inline void __bpf_generic_log_write(struct bpf_generic_log *log, const char *fmt,
>> + va_list args)
>> +{
>> + unsigned int n;
>> +
>> + n = vscnprintf(log->kbuf, BPF_GENERIC_TMP_LOG_SIZE, fmt, args);
>> +
>> + WARN_ONCE(n >= BPF_GENERIC_TMP_LOG_SIZE - 1,
>> + "bpf generic log truncated - local buffer too short\n");
>> +
>> + n = min(log->len_total - log->len_used - 1, n);
>> + log->kbuf[n] = '\0';
>> +
>> + if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
>> + log->len_used += n;
>> + else
>> + log->ubuf = NULL;
>> +}
>> +
>
> please see bpf_verifier_vlog() in kernel/bpf/log.c. We don't want to
> maintain another (even if light) version of it.
>
Agree with you, we should not maintain multiple versions of it.
However, it will be replaced with tracepoint patch.
Thanks,
Leon
