Re: scsi: Implement per-cpu logging buffer

[Hannes Reinecke]

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@xxxxxxxxxxxxxxx> wrote:
>>>> Gitweb:     http://git.kernel.org/linus/;a=commit;h=ded85c193a391a84076d5c6a7a5668fe164a490e
>>>> Commit:     ded85c193a391a84076d5c6a7a5668fe164a490e
>>>> Parent:     b0a93d96b2814c725161f91a4e35d0c29ec0f95b
>>>> Refname:    refs/heads/master
>>>> Author:     Hannes Reinecke <hare@xxxxxxx>
>>>> AuthorDate: Thu Jan 8 07:43:42 2015 +0100
>>>> Committer:  Christoph Hellwig <hch@xxxxxx>
>>>> 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@xxxxxx>
>>>>     Reviewed-by: Robert Elliott <elliott@xxxxxx>
>>>>     Signed-off-by: Hannes Reinecke <hare@xxxxxxx>
>>>>     Signed-off-by: Christoph Hellwig <hch@xxxxxx>
>>>
>>>> --- /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.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		               zSeries & Storage
hare@xxxxxxx			               +49 911 74053 688
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: F. Imendörffer, J. Smithard, J. Guild, D. Upmanyu, G. Norton
HRB 21284 (AG Nürnberg)
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/