[PATCH v4] iomap: support tail packing inline read

[Gao Xiang]

On Wed, Jul 21, 2021 at 08:43:00AM +0200, Andreas Grünbacher wrote:
> Am Mi., 21. Juli 2021 um 04:54 Uhr schrieb Gao Xiang
> <hsiangkao at linux.alibaba.com>:
> > Hi Andreas,
> >
> > On Wed, Jul 21, 2021 at 04:26:47AM +0200, Andreas Grünbacher wrote:
> > > Am Mi., 21. Juli 2021 um 02:33 Uhr schrieb Gao Xiang
> > > <hsiangkao at linux.alibaba.com>:
> > > > > And since you can only kmap one page at a time, an inline read grabs the
> > > > > first part of the data in "page one" and then we have to call
> > > > > iomap_begin a second time get a new address so that we can read the rest
> > > > > from "page two"?
> > > >
> > > > Nope, currently EROFS inline data won't cross page like this.
> > > >
> > > > But in principle, yes, I don't want to limit it to the current
> > > > EROFS or gfs2 usage. I think we could make this iomap function
> > > > more generally (I mean, I'd like to make the INLINE extent
> > > > functionity as general as possible,
> > >
> > > Nono. Can we please limit this patch what we actually need right now,
> > > and worry about extending it later?
> >
> > Can you elaborate what it will benefit us if we only support one tail
> > block for iomap_read_inline_data()? (I mean it has similar LOC changes,
> > similar implementation / complexity.) The only concern I think is if
> > it causes gfs2 regression, so that is what I'd like to confirm.
> 
> iomap_read_inline_data supports one inline page now (i.e., from the
> beginning of the file). It seems that you don't need more than one
> tail page in EROFS, right?
> 
> You were speculating about inline data in the middle of a file. That
> sounds like a really, really bad idea to me, and I don't think we
> should waste any time on it.

Huh? why do you think it's a bad idea? I could give real example to you.

At least, it can be used for some encoded data or repeated pattern (such
as AABBAABBAABB...) in a packed way (marked in extent metadata).
Is that enough?

Again, I don't see what the benefits if limiting it to one tail block,
it (maybe) just modifies:
+	/* handle tail-packing blocks cross the current page into the next */
+	size = min_t(unsigned int, iomap->length + pos - iomap->offset,
+		     PAGE_SIZE - poff);

to
+	size = min_t(unsigned int, i_blocksize(inode),
+		     PAGE_SIZE - poff);

And which has the similar complexity, so why not using iomap->length
here instead?

Thanks,
Gao Xiang