Re: [PATCH] net/core/net-procfs: use seq_put_decimal_ull_width() for decimal values in /proc/net/dev

[Paolo Abeni]

On 11/10/24 05:52, David Wang wrote:
> seq_printf() is costy, when reading /proc/net/dev, profiling indicates
> about 13% samples of seq_printf():
> 	dev_seq_show(98.350% 428046/435229)
> 	    dev_seq_printf_stats(99.777% 427092/428046)
> 		dev_get_stats(86.121% 367814/427092)
> 		    rtl8169_get_stats64(98.519% 362365/367814)
> 		    dev_fetch_sw_netstats(0.554% 2038/367814)
> 		    loopback_get_stats64(0.250% 919/367814)
> 		    dev_get_tstats64(0.077% 284/367814)
> 		    netdev_stats_to_stats64(0.051% 189/367814)
> 		    _find_next_bit(0.029% 106/367814)
> 		seq_printf(13.719% 58594/427092)
> And on a system with one wireless interface, timing for 1 million rounds of
> stress reading /proc/net/dev:
> 	real	0m51.828s
> 	user	0m0.225s
> 	sys	0m51.671s
> On average, reading /proc/net/dev takes ~0.051ms
> 
> With this patch, extra costs parsing format string by seq_printf() can be
> optimized out, and the timing for 1 million rounds of read is:
> 	real	0m49.127s
> 	user	0m0.295s
> 	sys	0m48.552s
> On average, ~0.048ms reading /proc/net/dev, a ~6% improvement.
> 
> Even though dev_get_stats() takes up the majority of the reading process,
> the improvement is still significant;
> And the improvement may vary with the physical interface on the system.
> 
> Signed-off-by: David Wang <00107082@xxxxxxx>

If the user-space is concerned with performances, it must use netlink.
Optimizing a legacy interface gives IMHO a very wrong message.

I'm sorry, I think we should not accept this change.

/P