Message-ID: <53F79FE8.7030406@interlog.com>
Date:	Fri, 22 Aug 2014 15:54:16 -0400
From:	Douglas Gilbert <dgilbert@...erlog.com>
To:	Yoshihiro YUNOMAE <yoshihiro.yunomae.ez@...achi.com>
CC:	Hannes Reinecke <hare@...e.de>,
	"Martin K. Petersen" <martin.petersen@...cle.com>,
	linux-scsi@...r.kernel.org, yrl.pp-manager.tt@...achi.com,
	linux-kernel@...r.kernel.org,
	"James E.J. Bottomley" <JBottomley@...allels.com>,
	Masami Hiramatsu <masami.hiramatsu.pt@...achi.com>,
	Hidehiro Kawai <hidehiro.kawai.ez@...achi.com>,
	Christoph Hellwig <hch@....de>
Subject: Re: [RFC PATCH -logging 00/10] scsi/constants: Output continuous
 error messages on trace

On 14-08-12 11:13 PM, Yoshihiro YUNOMAE wrote:
> Hi Douglas,
>
> Thank you for your comment.
>
> (2014/08/08 22:07), Douglas Gilbert wrote:
>> On 14-08-08 01:50 PM, Yoshihiro YUNOMAE wrote:
>>> Hi All,
>>>
>>> This patch set introduces new traceevents in order to output
>>> continuous error
>>> messages. Current SCSI printk messages in upstream kernel can be
>>> divided by and
>>> mixed with other messages. Even if each error message has its device id,
>>> sometimes we can easily be lost in mixed logs because the message's
>>> device id
>>> is separated from it's body. To avoid it, I'd like to use traceevents to
>>> store error messages into the ftrace or perf buuffer, because traceevents
>>> are atomically commited to the buffer.
>>>
>>> In this patch set, all printk messages are removed based on a local
>>> discussion with Hannes, but I think printk messages should be kept
>>> because all
>>> users don't enable traceevents and rsyslog can store as files.
>>>
>>> However, if printk of logging branch is kept, the messages are
>>> duplicate and
>>> it can induce stack overflow by using local buffer(*1).
>>>
>>>   (*1) https://lkml.org/lkml/2014/5/20/742
>>>
>>> So, my suggestion is follows:
>>>
>>> 1) printk
>>> Keeps current implemntation of upstream kernel.
>>> The messages are divided and can be mixed, but all users can check the
>>> error
>>> messages without any settings.
>>>
>>> 2) traceevents
>>> To get the complete messages, we can use ftrace or perf (or something
>>> on them).
>>> Users can always understand correct messages, but they need to set up the
>>> tracers.
>>>
>>> This patch set is based on Hannes' logging branch:
>>> http://git.kernel.org/cgit/linux/kernel/git/hare/scsi-devel.git/log/?h=logging
>>>
>>>
>>> [1/10] ~  [6/10]: just cleanup for logging branch
>>> [7/10] ~ [10/10]: introduce new traceevents
>>>
>>> Any comments are welcome!
>>
>> In sg3_utils there are now string yielding equivalents
>> to the sense buffer "print" functions. They take a form
>> like this:
>>    char * get_sense_str(const unsigned char * sense_buffer,
>>                         int sb_len, int blen, char * b);
>>
>> So this just does the hard work of decoding the sense buffer
>> (or saying it is invalid) the result of which it places in
>> buffer 'b'. And 'b' is returned (so this function can be in
>> the arguments of a driver's printing function).
>>
>> Adding such string functions would give other parts of the
>> SCSI subsystem the capability of tailoring their own
>> messages that include sense data information.
>>
>>
>> Existing sense buffer "print" function could be kept and
>> implemented using the newer "_str" variants. Would that
>> be worth the trouble?
>
> I have already sent the idea using local buffer on this February,
> but it was rejected by James (*1). By using stack region for local
> buffer, stack overflow can occur. So, I implemented this feature
> to atomically output an error message with device information.
>
> (*1) https://lkml.org/lkml/2014/5/20/742

Hi,
In the "_str" variants that I referred to, the caller provides
the buffer and its length. The responsibility of the
implementation of those "_str" variants is to use that
buffer, not exceed it, and make sure that it is null terminated
on return.

Can't see any inherent threat to the stack size there, and if
there is then that is just bad design by the caller.

The advantage of the "_str" variants is that the caller can
supply context and print/log a more useful message, perhaps
including the caller's __func__ . That message could include
sense information (perhaps truncated to 128 bytes, say),
and be output as a single unit.

IMO too many log messages are multi-line and in a noisy,
misbehaving system those messages can be interleaved
with other "noise" making them difficult to decipher.

Doug Gilbert


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists