Re: scsi: Implement per-cpu logging buffer

[Josh Triplett <josh@joshtriplett.org>]

On Fri, Feb 13, 2015 at 09:48:36AM +0100, Hannes Reinecke wrote:
> On 02/12/2015 06:18 PM, Josh Triplett wrote:
> > On Thu, Feb 12, 2015 at 02:29:35PM +0100, Hannes Reinecke wrote:
> >> On 02/12/2015 01:36 PM, Geert Uytterhoeven wrote:
> >>> On Wed, Feb 11, 2015 at 8:16 PM, Linux Kernel Mailing List
> >>> <linux-kernel@vger.kernel.org> wrote:
> >>>> Gitweb:     http://git.kernel.org/linus/;a=commit;h=ded85c193a391a84076d5c6a7a5668fe164a490e
> >>>> Commit:     ded85c193a391a84076d5c6a7a5668fe164a490e
> >>>> Parent:     b0a93d96b2814c725161f91a4e35d0c29ec0f95b
> >>>> Refname:    refs/heads/master
> >>>> Author:     Hannes Reinecke <hare@suse.de>
> >>>> AuthorDate: Thu Jan 8 07:43:42 2015 +0100
> >>>> Committer:  Christoph Hellwig <hch@lst.de>
> >>>> CommitDate: Fri Jan 9 15:44:28 2015 +0100
> >>>>
> >>>>     scsi: Implement per-cpu logging buffer
> >>>>
> >>>>     Implement a per-cpu buffer for formatting messages to avoid line breaks
> >>>>     up under high load.  This patch implements scmd_printk() and
> >>>>     sdev_prefix_printk() using the per-cpu buffer and makes sdev_printk() a
> >>>>     wrapper for sdev_prefix_printk().
> >>>>
> >>>>     Tested-by: Robert Elliott <elliott@hp.com>
> >>>>     Reviewed-by: Robert Elliott <elliott@hp.com>
> >>>>     Signed-off-by: Hannes Reinecke <hare@suse.de>
> >>>>     Signed-off-by: Christoph Hellwig <hch@lst.de>
> >>>
> >>>> --- /dev/null
> >>>> +++ b/drivers/scsi/scsi_logging.c
> >>>
> >>>> +#define SCSI_LOG_SPOOLSIZE 4096
> >>>> +#define SCSI_LOG_BUFSIZE 128
> >>>> +
> >>>> +#if (SCSI_LOG_SPOOLSIZE / SCSI_LOG_BUFSIZE) > BITS_PER_LONG
> >>>> +#warning SCSI logging bitmask too large
> >>>> +#endif
> >>>> +
> >>>> +struct scsi_log_buf {
> >>>> +       char buffer[SCSI_LOG_SPOOLSIZE];
> >>>> +       unsigned long map;
> >>>> +};
> >>>> +
> >>>> +static DEFINE_PER_CPU(struct scsi_log_buf, scsi_format_log);
> >>>
> >>> Do we really need a static 4 KiB per-CPU buffer?
> >>>
> >>> bloat-o-meter:
> >>>
> >>> add/remove: 183/94 grow/shrink: 314/211 up/down: 33467/-21291 (12176)
> >>> function                                     old     new   delta
> >>> scsi_format_log                                -    4100   +4100
> >>> handle_mm_fault                             1794    2750    +956
> >>> scsi_log_print_sense_hdr                       -     774    +774
> >>> proc_keys_show                                 -     770    +770
> >>>
> >> Define 'need'.
> >> We don't absolutely 'need' it. (Configure it out and it's gone).
> >>
> >> But when we want to avoid several logging messages coming in from
> >> various CPUs overwriting each other and _not_ introduce additional
> >> latency by locking a single buffer, then yes.
> >>
> >> We can possibly reduce it to, say, 1KiB or even lower by imposing
> >> stricter rules on the logging functions.
> >> But I don't see a way around the per-CPU buffer.
> > 
> > It seems very odd to introduce a mechanism like this specifically for
> > SCSI, rather than introducing a generic per-CPU buffered-print mechanism
> > in printk, controlled by a config option.  That option could then
> > automatically go away when !SMP, or !PRINTK, or if users don't actually
> > care about message ordering.
> > 
> But then we ran afoul with the printk purists.
> 
> Thing is, if we were to use per-CPU buffers for printk() out of
> necessity we have to queue these buffers for writing out.
> So there is a time window during which the message already is in the
> per-CPU buffer but still not printed out as printk() is currently
> writing out one of the other per-CPU buffers.
> 
> If there is a consensus that such a delayed printk() is useful and a
> valid use case then yes, sure I can give it a go.
> 
> Personally I think printk() currently has an unfortunate double
> purpose: on the one hand it should print out emergency messages
> immediate so that they'll be visible if the system crashes. On the
> other hand it is used as a general logging facility, where frankly
> most of the subsystems simple do not care at all if and when the
> message are printed.
> Splitting that off would indeed be a good idea, as then we can have
> the ultra-fast, go to console now printk() thingie, and another 'hey
> I don't care, just wanted to let you know something happened'
> delayed logging output.
> 
> But I certainly will not attempt to implement this without a broader
> consensus. Typically patching printk is a good way of getting flamed.

I'm not suggesting that you change printk(); I'm just suggesting that
you drop the scsi_* prefixes from the buffered logging mechanism you've
already created, and make it a generic buffered logging mechanism that
other subsystems can use if desired.  More importantly, though, whether
you keep it SCSI-specific or not, please consider making it
configurable, dependent on SMP and PRINTK, and fall through to just
printk if configured out.

- Josh Triplett