Re: [PATCH] lib/core/coio_file: Use eio_sendfile_sync instead of a chunk mode

[Vladimir Davydov <vdavydov.dev@gmail.com>]

On Fri, Apr 12, 2019 at 05:05:58PM +0300, Cyrill Gorcunov wrote:
> On Fri, Apr 12, 2019 at 04:43:23PM +0300, Vladimir Davydov wrote:
> > > +	ssize_t nwritten = eio_sendfile_sync(dest_fd, source_fd, 0, st.st_size);
> > > +	if (nwritten != st.st_size) {
> > > +		say_syserror("sendfile, [%s -> %s]",
> > > +			     fio_filename(source_fd),
> > > +			     fio_filename(dest_fd));
> > 
> > Other coio/fio methods don't log errors so I don't think this one
> > should, either.
> 
> But fio_ methods which we used before the patch do print an error,
> so we should continue printings.

Ah, okay, I see.

> 
> > An extract from the sendfile manual page:
> > 
> > } sendfile() will transfer at most 0x7ffff000 (2,147,479,552) bytes,
> > } returning the number of bytes actually transferred. (This is true on
> > } both 32-bit and 64-bit systems.)
> > 
> > So I don't think using sendfile is equivalent to read+write...
> 
> Actaully man page is a bit wrong, at most it may handle up to 0x7fffffff
> bytes. And here is a question do we really ever going to transfer such big
> files? Moreover using read/write with 4K buffer for big files is pure and
> utter crap, the only thing which saves situation a bit is a page cache,
> but still such small buffer requires a number of context switches :)

It's not a very hot function, to begin with. I doubt it's used that
often or for copying big files. Besides, it is handled in a coio thread
(i.e. not in tx) so we can afford it to be a bit slow.

> 
> But I agree that using sendfile here actually may trim file. I'll take
> a look at evening but I think we should rather use sendfile with offsets
> in a cycle. Seriouly, using 4K buffer is the worst possible algo :)

Yeah, it would be great to make use of sendfile, but we must make sure
we won't occasionally break anything.