lkml.org 
[lkml]   [2018]   [Sep]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH 2/6] pstore: Add event tracing support
    On Tue, 18 Sep 2018 23:22:48 +0530
    Sai Prakash Ranjan <saiprakash.ranjan@codeaurora.org> wrote:

    > On 9/18/2018 5:04 AM, Steven Rostedt wrote:
    > >
    > > It looks like pstore_event_call() gets called from a trace event. You
    > > can't call kmalloc() from one. One thing is that kmalloc has
    > > tracepoints itself. You trace those you just entered an infinite loop.
    > >
    > >
    >
    > Ok will remove it in v2. But any alternative way to do this?

    I think I describe it below.

    >
    > >> +
    > >> + event_call = fbuffer->trace_file->event_call;
    > >> + if (!event_call || !event_call->event.funcs ||
    > >> + !event_call->event.funcs->trace)
    > >> + goto fail_event;
    > >> +
    > >> + event = &fbuffer->trace_file->event_call->event;
    > >> +
    > >> + spin_lock_irqsave(&psinfo->buf_lock, flags);
    > >> +
    > >> + trace_seq_init(&iter->seq);
    > >> + iter->ent = fbuffer->entry;
    > >
    > > I guess what you are doing is needing to translate the raw data into
    > > ascii output, and need the trace_iterator to do so.
    > >
    > > You are already under a psinfo->buf_lock. Add a dummy iterator to that
    > > and use it instead.
    > >
    > > trace_seq_init(&psinfo->iter->seq);
    > >
    > >> + event_call->event.funcs->trace(iter, 0, event);
    > >
    > > (psinfo->iter, 0 , event);
    > >
    > > etc.
    > >
    >
    > Sure, will update in v2.
    >
    > >> + trace_seq_putc(&iter->seq, 0);
    > >> +
    > >> + if (seq->size > psinfo->bufsize)
    > >> + seq->size = psinfo->bufsize;
    > >> +
    > >> + s = &iter->seq;
    > >> + seq = &s->seq;
    > >> +
    > >> + record.buf = (char *)(seq->buffer);
    > >> + record.size = seq->len;
    > >> + psinfo->write(&record);
    > >> +
    > >> + spin_unlock_irqrestore(&psinfo->buf_lock, flags);
    > >
    > > You may also need to convert these spin_locks into raw_spin_locks as
    > > when PREEMPT_RT enters the kernel you don't want them to turn into
    > > mutexes.
    > >
    > > But that can be another patch.
    > >
    >
    > I will change this in v2, but can't we have it in same patch?

    I suggested a separate patch because buf_lock is used elsewhere.
    Changing it to "raw_spin_lock" will affect more than just what this
    patch series does. Thus, I recommend making it a separate patch to keep
    this patch series from being more intrusive than it needs to be.

    Thanks!

    -- Steve

    \
     
     \ /
      Last update: 2018-09-18 22:45    [W:4.605 / U:0.012 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site