[PATCH v3] cxl: Export AFU error buffer via sysfs

[Michael Neuling]

On Thu, 2015-05-21 at 09:06 +0530, trigg wrote:
> 
> 
> > On 21-May-2015, at 05:16, Michael Neuling <mikey at neuling.org> wrote:
> > 
> > + */
> > > +ssize_t cxl_afu_read_err_buffer(struct cxl_afu *afu, char *buf,
> > > +                loff_t off, size_t count)
> > > +{
> > > +    loff_t aligned_off;
> > > +    size_t aligned_count;
> > > +    const void __iomem *ebuf = afu->afu_desc_mmio +
> > > afu->eb_offset;
> > > +
> > > +    if (!afu->eb_len || count == 0 || off < 0)
> > 
> > if eb_len == 0 we don't even create the sysfs file.  So is this
> > check
> > needed?
> This function is non static so it can be potentially called from
> kernel.

What?  You mean outside this file?

> Condition check of "if (count == 0 || off < 0 || (size_t)off
> >= afu->eb_len) "
> should work solving the problem below too.

It makes no sense to call this outside of sysfs reading the error
buffer.

> > > +    aligned_off = round_down(off, 8);
> > > +    aligned_count = round_up(off + count, 8) - aligned_off;
> > 
> > I kinda preferred the start end variables, and length was just end -
> > start.
> > 
> > I though it was more readable. IMHO
> > 
> > How about:
> > 
> >   aligned_start = round_down(off, 8);
> >   aligned_end = round_up(off + count, 8);
> >   aligned_length = aligned_end - aligned_start;
> Agreed on aligned_start and aligned_end but would not
> need aligned_end in rest of the code.

Aligned_end makes it more readable.

> > 
> > > +
> > > +    /* fast path */
> > > +    if ((aligned_off == off) && (aligned_count == count)) {
> > > +        /* no need to use the bounce buffer */
> > > +        _memcpy_fromio(buf, ebuf + off, count);
> > 
> > I would drop this, as the other code path should work fine.
> > Premature optimisation.....
> I am inclined to differ on this. Code below uses a bounce buffer
> which needs more than double the amount of loads and stores.

Cacheable vs non-cacheable is important here though.

> If the caller wants to speed up things it can simply ask for aligned
> read that won't have this overhead. This will be especially useful 
> In large reads.

The overhead will mostly be in the non-cachable/MMIO load/stores rather
than the cacheable load/stores, so there may be little performance
difference  anyway.

The only reason this could be useful is if you have a really really
large error buffer.  Even then, why do you need to read it that quickly?

Mikey