RE: [PATCH] PM / devfreq: Convert to seq_buf

[Biju Das]

Hi Geert,

Thanks for the feedback.

> -----Original Message-----
> From: Geert Uytterhoeven <geert@xxxxxxxxxxxxxx>
> Sent: 07 August 2026 13:19
> Subject: Re: [PATCH] PM / devfreq: Convert to seq_buf
> 
> Hi Biju,
> 
> On Fri, 7 Aug 2026 at 12:03, Biju <biju.das.au@xxxxxxxxx> wrote:
> > From: Biju Das <biju.das.jz@xxxxxxxxxxxxxx>
> >
> > The devfreq transition table sysfs attribute (trans_stat_show())
> > builds its output using manual sysfs_emit_at() calls with hand-rolled
> > length tracking and PAGE_SIZE bounds checks scattered throughout the loops.
> >
> > On SoCs with large frequency tables, such as Renesas RZ/G3L, the
> > resulting table can exceed PAGE_SIZE. In that case the function bails
> > out entirely and returns -EFBIG, logging:
> >
> >   devfreq transition table exceeds PAGE_SIZE. Disabling
> >
> > which makes the attribute completely unusable on affected platforms.
> >
> > Convert trans_stat_show() to use struct seq_buf instead.
> > seq_buf_printf() tracks the buffer position and handles overflow
> > internally, removing the need for manual length checks before every
> > write. If the table still overflows PAGE_SIZE, truncate the output to
> > PAGE_SIZE - 1, NUL-terminate it, and log a rate-limited warning rather
> > than failing the read outright, so the attribute still returns a
> > usable (if truncated) table instead of nothing.
> >
> > Suggested-by: Geert Uytterhoeven <geert+renesas@xxxxxxxxx>
> > Signed-off-by: Biju Das <biju.das.jz@xxxxxxxxxxxxxx>
> 
> Thanks for your patch!
> 
> > --- a/drivers/devfreq/devfreq.c
> > +++ b/drivers/devfreq/devfreq.c
> 
> > @@ -1661,9 +1662,10 @@ static ssize_t trans_stat_show(struct device
> > *dev,
> 
> > +       /*
> > +        * If an overflow occurs, seq_buf_used() can return a value greater
> > +        * than PAGE_SIZE. Clamp the return length to PAGE_SIZE - 1 and
> > +        * ensure a null-terminator is written to prevent the sysfs
> > +        * "bad count" or out-of-bounds warning.
> > +        */
> > +       if (seq_buf_has_overflowed(&s)) {
> > +               pr_warn_ratelimited("devfreq transition table truncated due to PAGE_SIZE limit\n");
> > +               len = PAGE_SIZE - 1;
> > +               buf[len] = '\0';
> 
> So shouldn't you use seq_file instead of seq_buf, to lift this limitation?
> devfreq_summary_show() already uses that.

I have explored this option and got this response from AI.

Here it is sysfs, not debugfs. 

The limitation comes from sysfs's one-page buffer contract
not from the choice of seq_buf/seq_file. To actually lift the limitation,
this attribute would need to move off sysfs device_attribute
entirely.

The problem is that trans_stat_show() is a sysfs device_attribute show() callback,
not a debugfs file opened through single_open()/seq_read(). Sysfs attribute reads
are fundamentally capped at PAGE_SIZE — the buf passed in is a single, fixed-size
page, and there's no .read()/.open() VFS machinery driving repeated .show() calls
with buffer growth. So even switching to struct seq_file, we'd have to hand-initialize
it over that same fixed PAGE_SIZE buffer.

I am agreeing to AI response. Please correct me, if it is not the case.

Cheers,
Biju