On Thu, Jan 11, 2024 at 07:25:13PM +0000, Matthew Wilcox (Oracle) wrote:
> Modelled after the loop in iomap_write_iter(), copy larger chunks from
> userspace if the filesystem has created large folios.
Heh, I hadn't tested this on ext4, and it blows up spectacularly.
We hit the first of these two BUG_ONs:
+++ b/fs/buffer.c
@@ -2078,8 +2078,8 @@ int __block_write_begin_int(struct folio *folio, loff_t pos, unsigned len,
struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
BUG_ON(!folio_test_locked(folio));
- BUG_ON(to > folio_size(folio));
- BUG_ON(from > to);
+ if (from < to || to > folio_size(folio))
+ to = folio_size(folio);
head = folio_create_buffers(folio, inode, 0);
blocksize = head->b_size;
I'm just doing an xfstests run now, but I think this will fix the problem
(it's up to 400 seconds after crashing at 6 seconds the first time).
Essentially if we pass in a length larger than the folio we find, we
just clamp the amount of work we do to the size of the folio.
> Signed-off-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> ---
> mm/filemap.c | 40 +++++++++++++++++++++++++---------------
> 1 file changed, 25 insertions(+), 15 deletions(-)
>
> diff --git a/mm/filemap.c b/mm/filemap.c
> index 750e779c23db..964d5d7b9721 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -3893,6 +3893,7 @@ EXPORT_SYMBOL(generic_file_direct_write);
> ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i)
> {
> struct file *file = iocb->ki_filp;
> + size_t chunk = PAGE_SIZE << MAX_PAGECACHE_ORDER;
> loff_t pos = iocb->ki_pos;
> struct address_space *mapping = file->f_mapping;
> const struct address_space_operations *a_ops = mapping->a_ops;
> @@ -3901,16 +3902,18 @@ ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i)
>
> do {
> struct page *page;
> - unsigned long offset; /* Offset into pagecache page */
> - unsigned long bytes; /* Bytes to write to page */
> + struct folio *folio;
> + size_t offset; /* Offset into folio */
> + size_t bytes; /* Bytes to write to folio */
> size_t copied; /* Bytes copied from user */
> void *fsdata = NULL;
>
> - offset = (pos & (PAGE_SIZE - 1));
> - bytes = min_t(unsigned long, PAGE_SIZE - offset,
> - iov_iter_count(i));
> + bytes = iov_iter_count(i);
> +retry:
> + offset = pos & (chunk - 1);
> + bytes = min(chunk - offset, bytes);
> + balance_dirty_pages_ratelimited(mapping);
>
> -again:
> /*
> * Bring in the user page that we will copy from _first_.
> * Otherwise there's a nasty deadlock on copying from the
> @@ -3932,11 +3935,16 @@ ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i)
> if (unlikely(status < 0))
> break;
>
> + folio = page_folio(page);
> + offset = offset_in_folio(folio, pos);
> + if (bytes > folio_size(folio) - offset)
> + bytes = folio_size(folio) - offset;
> +
> if (mapping_writably_mapped(mapping))
> - flush_dcache_page(page);
> + flush_dcache_folio(folio);
>
> - copied = copy_page_from_iter_atomic(page, offset, bytes, i);
> - flush_dcache_page(page);
> + copied = copy_folio_from_iter_atomic(folio, offset, bytes, i);
> + flush_dcache_folio(folio);
>
> status = a_ops->write_end(file, mapping, pos, bytes, copied,
> page, fsdata);
> @@ -3954,14 +3962,16 @@ ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i)
> * halfway through, might be a race with munmap,
> * might be severe memory pressure.
> */
> - if (copied)
> + if (chunk > PAGE_SIZE)
> + chunk /= 2;
> + if (copied) {
> bytes = copied;
> - goto again;
> + goto retry;
> + }
> + } else {
> + pos += status;
> + written += status;
> }
> - pos += status;
> - written += status;
> -
> - balance_dirty_pages_ratelimited(mapping);
> } while (iov_iter_count(i));
>
> if (!written)
> --
> 2.43.0
>
