by Brian Foster

[permalink] [raw]

Subject: Re: [PATCH 19/19] writeback: simplify writeback iteration

On Wed, Jan 31, 2024 at 08:14:37AM +0100, Christoph Hellwig wrote:
> On Tue, Jan 30, 2024 at 10:50:16PM +0100, Jan Kara wrote:
> > Well, batch release needs to be only here because if writeback_get_folio()
> > returns NULL, the batch has been already released by it.
>
> Indeed.
>
> > So what would be
> > duplicated is only the error assignment. But I'm fine with the version in
> > the following email and actually somewhat prefer it compared the yet
> > another variant you've sent.
>
> So how about another variant, this is closer to your original suggestion.
> But I've switched around the ordered of the folio or not branches
> from my original patch, and completely reworked and (IMHO) improved the
> comments. it replaces patch 19 instead of being incremental to be
> readable:
>

Not a thorough review but at first glance I like it. I think the direct
function call makes things easier to follow. Just one quick thought...

..
> diff --git a/mm/page-writeback.c b/mm/page-writeback.c
> index 0763c4353a676a..eefcb00cb7b227 100644
> --- a/mm/page-writeback.c
> +++ b/mm/page-writeback.c
..
> @@ -2458,60 +2433,107 @@ static struct folio *writeback_get_folio(struct address_space *mapping,
> return folio;
> }
>
..
> +struct folio *writeback_iter(struct address_space *mapping,
> + struct writeback_control *wbc, struct folio *folio, int *error)
> {
> - if (wbc->range_cyclic)
> - wbc->index = mapping->writeback_index; /* prev offset */
> - else
> - wbc->index = wbc->range_start >> PAGE_SHIFT;
> -
> - if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
> - tag_pages_for_writeback(mapping, wbc->index, wbc_end(wbc));
> -
> - wbc->err = 0;
> - folio_batch_init(&wbc->fbatch);
> - return writeback_get_folio(mapping, wbc);
> -}
> + if (!folio) {
> + folio_batch_init(&wbc->fbatch);
> + wbc->err = 0;

The implied field initialization via !folio feels a little wonky to me
just because it's not clear from the client code that both fields must
be initialized. Even though the interface is simpler, I wonder if it's
still worth having a dumb/macro type init function that at least does
the batch and error field initialization.

Or on second thought maybe having writeback_iter() reset *error as well
on folio == NULL might be a little cleaner without changing the
interface....

Brian

>
> -struct folio *writeback_iter_next(struct address_space *mapping,
> - struct writeback_control *wbc, struct folio *folio, int error)
> -{
> - unsigned long nr = folio_nr_pages(folio);
> + /*
> + * For range cyclic writeback we remember where we stopped so
> + * that we can continue where we stopped.
> + *
> + * For non-cyclic writeback we always start at the beginning of
> + * the passed in range.
> + */
> + if (wbc->range_cyclic)
> + wbc->index = mapping->writeback_index;
> + else
> + wbc->index = wbc->range_start >> PAGE_SHIFT;
>
> - wbc->nr_to_write -= nr;
> + /*
> + * To avoid livelocks when other processes dirty new pages, we
> + * first tag pages which should be written back and only then
> + * start writing them.
> + *
> + * For data-integrity sync we have to be careful so that we do
> + * not miss some pages (e.g., because some other process has
> + * cleared the TOWRITE tag we set). The rule we follow is that
> + * TOWRITE tag can be cleared only by the process clearing the
> + * DIRTY tag (and submitting the page for I/O).
> + */
> + if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
> + tag_pages_for_writeback(mapping, wbc->index,
> + wbc_end(wbc));
> + } else {
> + wbc->nr_to_write -= folio_nr_pages(folio);
>
> - /*
> - * Handle the legacy AOP_WRITEPAGE_ACTIVATE magic return value.
> - * Eventually all instances should just unlock the folio themselves and
> - * return 0;
> - */
> - if (error == AOP_WRITEPAGE_ACTIVATE) {
> - folio_unlock(folio);
> - error = 0;
> + /*
> + * For integrity writeback we have to keep going until we have
> + * written all the folios we tagged for writeback prior to
> + * entering the writeback loop, even if we run past
> + * wbc->nr_to_write or encounter errors, and just stash away
> + * the first error we encounter in wbc->err so that it can
> + * be retrieved on return.
> + *
> + * This is because the file system may still have state to clear
> + * for each folio. We'll eventually return the first error
> + * encountered.
> + */
> + if (wbc->sync_mode == WB_SYNC_ALL) {
> + if (*error && !wbc->err)
> + wbc->err = *error;
> + } else {
> + if (*error || wbc->nr_to_write <= 0)
> + goto done;
> + }
> }
>
> - if (error && !wbc->err)
> - wbc->err = error;
> + folio = writeback_get_folio(mapping, wbc);
> + if (!folio) {
> + /*
> + * For range cyclic writeback not finding another folios means
> + * that we are at the end of the file. In that case go back
> + * to the start of the file for the next call.
> + */
> + if (wbc->range_cyclic)
> + mapping->writeback_index = 0;
>
> - /*
> - * For integrity sync we have to keep going until we have written all
> - * the folios we tagged for writeback prior to entering the writeback
> - * loop, even if we run past wbc->nr_to_write or encounter errors.
> - * This is because the file system may still have state to clear for
> - * each folio. We'll eventually return the first error encountered.
> - *
> - * For background writeback just push done_index past this folio so that
> - * we can just restart where we left off and media errors won't choke
> - * writeout for the entire file.
> - */
> - if (wbc->sync_mode == WB_SYNC_NONE &&
> - (wbc->err || wbc->nr_to_write <= 0)) {
> - writeback_finish(mapping, wbc, folio->index + nr);
> - return NULL;
> + /*
> + * Return the first error we encountered (if there was any) to
> + * the caller now that we are done.
> + */
> + *error = wbc->err;
> }
> + return folio;
>
> - return writeback_get_folio(mapping, wbc);
> +done:
> + if (wbc->range_cyclic)
> + mapping->writeback_index = folio->index + folio_nr_pages(folio);
> + folio_batch_release(&wbc->fbatch);
> + return NULL;
> }
>
> /**
> @@ -2549,13 +2571,18 @@ int write_cache_pages(struct address_space *mapping,
> struct writeback_control *wbc, writepage_t writepage,
> void *data)
> {
> - struct folio *folio;
> - int error;
> + struct folio *folio = NULL;
> + int error = 0;
>
> - for_each_writeback_folio(mapping, wbc, folio, error)
> + while ((folio = writeback_iter(mapping, wbc, folio, &error))) {
> error = writepage(folio, wbc, data);
> + if (error == AOP_WRITEPAGE_ACTIVATE) {
> + folio_unlock(folio);
> + error = 0;
> + }
> + }
>
> - return wbc->err;
> + return error;
> }
> EXPORT_SYMBOL(write_cache_pages);
>
> @@ -2563,13 +2590,17 @@ static int writeback_use_writepage(struct address_space *mapping,
> struct writeback_control *wbc)
> {
> struct blk_plug plug;
> - struct folio *folio;
> - int err;
> + struct folio *folio = NULL;
> + int err = 0;
>
> blk_start_plug(&plug);
> - for_each_writeback_folio(mapping, wbc, folio, err) {
> + while ((folio = writeback_iter(mapping, wbc, folio, &err))) {
> err = mapping->a_ops->writepage(&folio->page, wbc);
> mapping_set_error(mapping, err);
> + if (err == AOP_WRITEPAGE_ACTIVATE) {
> + folio_unlock(folio);
> + err = 0;
> + }
> }
> blk_finish_plug(&plug);
>
>

2024-01-31 13:35:04

by Christoph Hellwig

[permalink] [raw]

Subject: Re: [PATCH 19/19] writeback: simplify writeback iteration

On Wed, Jan 31, 2024 at 07:50:25AM -0500, Brian Foster wrote:
> The implied field initialization via !folio feels a little wonky to me
> just because it's not clear from the client code that both fields must
> be initialized. Even though the interface is simpler, I wonder if it's
> still worth having a dumb/macro type init function that at least does
> the batch and error field initialization.
>
> Or on second thought maybe having writeback_iter() reset *error as well
> on folio == NULL might be a little cleaner without changing the
> interface....

I like that second idea. An initialization helper I could live with,
but if only folio needs a defined state, that seems superflous.