On Tue, Sep 25, 2018 at 1:37 PM, Joel Fernandes <joelaf@xxxxxxxxxx> wrote: > On Sun, Sep 23, 2018 at 8:33 AM, Sai Prakash Ranjan > <saiprakash.ranjan@xxxxxxxxxxxxxx> wrote: >> On 9/22/2018 10:07 PM, Sai Prakash Ranjan wrote: >>> >>> On 9/22/2018 2:35 PM, Joel Fernandes wrote: >>>> >>>> On Sat, Sep 8, 2018 at 4:28 PM Sai Prakash Ranjan >>>> <saiprakash.ranjan@xxxxxxxxxxxxxx> wrote: >>>>> >>>>> >>>>> + >>>>> + trace_seq_init(&iter->seq); >>>>> + iter->ent = fbuffer->entry; >>>>> + event_call->event.funcs->trace(iter, 0, event); >>>>> + trace_seq_putc(&iter->seq, 0); >>>> >>>> >>>> Would it be possible to store the binary trace record in the pstore >>>> buffer instead of outputting text? I suspect that will both be faster >>>> and less space. >>>> >>> >>> I will try this and come back. >>> >> >> Hi Joel, >> >> I removed trace_seq_putc and there is some improvement seen: 203 MB/s >> >> # dd if=/dev/zero of=/dev/null status=progress >> 12207371264 bytes (12 GB, 11 GiB) copied, 60 s, 203 MB/s^C >> 24171926+0 records in >> 24171926+0 records out >> 12376026112 bytes (12 GB, 12 GiB) copied, 60.8282 s, 203 MB/s >> >> This seems good when compared to 190 MB/s seen previously. >> If this is Ok, then I will spin v2 with changes suggested. > > Sorry for slow reply, yes that sounds good and a worthwhile perf improvement. > Well so I think you should still not use spinlock to synchronize and split the buffer. You could expand pstore_record to have a ts field or introduce a new API like ->write_percpu instead of write, or something. But I strongly feel you should lock. For ftrace function trace, the perf reduction with locking was dramatic. - Joel