I have Linux 5.12-rc4.
On top of that I have "David's patches", an 85 patch
series I peeled off of David Howell's fscache-iter
branch on git.kernel.org a few days ago. These
patches include what I need to use
readahead_expand.
On top of that is my orangefs_readahead patch.
I have run through the xfstests I run, and am
failing generic 75 112 127 & 263, which I
normally pass.
So I got rid of my orangefs_readahead patch.
Still failing 75 112 127 & 263.
Then I got rid of David's patches, I'm at
generic Linux 5.12-rc4, and am no longer
failing those tests.
Just thought I should mention it... other
than that, I'm real happy with the
orangefs_readahead patch, it is a
giant improvement. Y'all's help
made all the difference...
-Mike
On Sat, Mar 27, 2021 at 11:04 PM Mike Marshall <hubcap@xxxxxxxxxxxx> wrote:
>
> This seems OK... ?
>
> static void orangefs_readahead(struct readahead_control *rac)
> {
> loff_t offset;
> struct iov_iter iter;
> struct file *file = rac->file;
> struct inode *inode = file->f_mapping->host;
> struct xarray *i_pages;
> struct page *page;
> loff_t new_start = readahead_pos(rac);
> int ret;
> size_t new_len = 0;
>
> loff_t bytes_remaining = inode->i_size - readahead_pos(rac);
> loff_t pages_remaining = bytes_remaining / PAGE_SIZE;
>
> if (pages_remaining >= 1024)
> new_len = 4194304;
> else if (pages_remaining > readahead_count(rac))
> new_len = bytes_remaining;
>
> if (new_len)
> readahead_expand(rac, new_start, new_len);
>
> offset = readahead_pos(rac);
> i_pages = &file->f_mapping->i_pages;
>
> iov_iter_xarray(&iter, READ, i_pages, offset, readahead_length(rac));
>
> /* read in the pages. */
> if ((ret = wait_for_direct_io(ORANGEFS_IO_READ, inode,
> &offset, &iter, readahead_length(rac),
> inode->i_size, NULL, NULL, file)) < 0)
> gossip_debug(GOSSIP_FILE_DEBUG,
> "%s: wait_for_direct_io failed. \n", __func__);
> else
> ret = 0;
>
> /* clean up. */
> while ((page = readahead_page(rac))) {
> page_endio(page, false, ret);
> put_page(page);
> }
> }
>
> I need to go remember how to git send-email through the
> kernel.org email server, I apologize for the way gmail
> unformats my code, even in plain text mode...
>
> -Mike
>
> On Sat, Mar 27, 2021 at 11:56 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >
> > On Sat, Mar 27, 2021 at 11:40:08AM -0400, Mike Marshall wrote:
> > > int ret;
> > >
> > > loff_t new_start = readahead_index(rac) * PAGE_SIZE;
> >
> > That looks like readahead_pos() to me.
> >
> > > size_t new_len = 524288;
> > > readahead_expand(rac, new_start, new_len);
> > >
> > > npages = readahead_count(rac);
> > > offset = readahead_pos(rac);
> > > i_pages = &file->f_mapping->i_pages;
> > >
> > > iov_iter_xarray(&iter, READ, i_pages, offset, npages * PAGE_SIZE);
> >
> > readahead_length()?
> >
> > > /* read in the pages. */
> > > ret = wait_for_direct_io(ORANGEFS_IO_READ, inode, &offset, &iter,
> > > npages * PAGE_SIZE, inode->i_size, NULL, NULL, file);
> > >
> > > /* clean up. */
> > > while ((page = readahead_page(rac))) {
> > > page_endio(page, false, 0);
> > > put_page(page);
> > > }
> > > }
> >
> > What if wait_for_direct_io() returns an error? Shouldn't you be calling
> >
> > page_endio(page, false, ret)
> >
> > ?
> >
> > > On Sat, Mar 27, 2021 at 9:57 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > > >
> > > > On Sat, Mar 27, 2021 at 08:31:38AM +0000, David Howells wrote:
> > > > > However, in Mike's orangefs_readahead_cleanup(), he could replace:
> > > > >
> > > > > rcu_read_lock();
> > > > > xas_for_each(&xas, page, last) {
> > > > > page_endio(page, false, 0);
> > > > > put_page(page);
> > > > > }
> > > > > rcu_read_unlock();
> > > > >
> > > > > with:
> > > > >
> > > > > while ((page = readahead_page(ractl))) {
> > > > > page_endio(page, false, 0);
> > > > > put_page(page);
> > > > > }
> > > > >
> > > > > maybe?
> > > >
> > > > I'd rather see that than open-coded use of the XArray. It's mildly
> > > > slower, but given that we're talking about doing I/O, probably not enough
> > > > to care about.
