From: Greg KH <gregkh@suse.de>
Subject: Re: RFC: seq_file: optimize memory re-alloc strategy for seq file
 sizes > PAGE_SIZE
Date: Fri, 15 Jul 2011 17:18:43 +0200
Message-ID: <20110715151843.GB4747@suse.de>
References: <4E1FFF94.4080803@st.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Cc: linux-fsdevel@vger.kernel.org,
	Stuart MENEFY <stuart.menefy@st.com>,
	srinivas.kandagatla@gmail.com
To: Srinivas KANDAGATLA <srinivas.kandagatla@st.com>
Content-Disposition: inline
In-Reply-To: <4E1FFF94.4080803@st.com>
Sender: linux-fsdevel-owner@vger.kernel.org

On Fri, Jul 15, 2011 at 09:51:32AM +0100, Srinivas KANDAGATLA wrote:
> Hi All,
> I have noticed that, reading or dumping a full sysfs entry (ex:
> /sys/kernel/debug/gpio )

That's not sysfs, it's debugfs.

> which exceeds PAGE_SIZE restarts associated seq_file iterator
> (calling start->show->stop)
> morethan one time.
> 
> In my case gpio entry just exceeded 8k which resulted in show
> function getting
> called 3 times.
> 
> The reason is because the allocated buffer overflows and seq_file allocates
> new buffer of next order and starts iteration from the start, which
> is clear in
> seq_file code.
> However this is bit of conservative approach which results in:
> - restarting iterator if file sz exceeds PAGE_SIZE( which is really
> not necessary )
> - not reusing the formated buffer (which can be done by re-alloc and copy).
> - Will help entires like dumping cacheline or pagetables to capture
> its state with better accuracy.
> 
> To address this situation, I think seq_file formating/printf
> functions should
> re-allocate buffer as soon as they detect buffer overflow.

This is debugging code, why should performance matter?

> This patch attempts to do the same by adding new function
> buf_realloc which inturn
> calls krealloc and all the seq_file formating or printing functions
> call this
> function whenever an overflow is detected and retry.
> 
> 
> Main motive of this patch is to avoid restarting the iterator and NOT
> performance, however there is a bit of gain in performance too.
> 
> Performance figures(average over 100 times loop) for different sizes of
> sysfs entries:
> 
> without realloc:
> ---------------------------------
>   SZ        REAL    USER    SYS
> ---------------------------------
> | 4k   | 0.012    0.002    0.009
> | 8k   | 0.013    0.002    0.010
> | 16k  | 0.015    0.003    0.012
> | 32k  | 0.019    0.002    0.017
> --------------------------------
> 
> with realloc:
> ---------------------------------
>   SZ        REAL    USER    SYS
> ---------------------------------
> | 4k  |    0.012    0.002    0.009
> | 8k  |    0.012    0.002    0.009
> | 16k |    0.013    0.002    0.010
> | 32k |    0.015    0.003    0.012
> -------------------------------
> 
> 
> I discovered this issue while I was working on a different issue as
> I happed
> to stick a printk in show function, which got called 3 times when I
> dumped the
> entry with 'cat' :-). which I did not expect.
> 
> 
> comments? :-)

Again, debugfs doesn't matter for this type of thing, so adding
complexity to the core kernel code for it shouldn't be done unless
absolutely needed.

thanks,

greg k-h