2024-02-26 20:25:28

by Matthew Wilcox

Subject: Re: [PATCH 10/13] iomap: fix iomap_dio_zero() for fs bs > system page size

On Mon, Feb 26, 2024 at 10:49:33AM +0100, Pankaj Raghav (Samsung) wrote:
> +++ b/fs/iomap/direct-io.c
> @@ -239,14 +239,23 @@ static void iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
> struct page *page = ZERO_PAGE(0);
> struct bio *bio;
>
> - bio = iomap_dio_alloc_bio(iter, dio, 1, REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
> + WARN_ON_ONCE(len > (BIO_MAX_VECS * PAGE_SIZE));
> +
> + bio = iomap_dio_alloc_bio(iter, dio, BIO_MAX_VECS,
> + REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
> fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
> GFP_KERNEL);
> +
> bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
> bio->bi_private = dio;
> bio->bi_end_io = iomap_dio_bio_end_io;
>
> - __bio_add_page(bio, page, len, 0);
> + while (len) {
> + unsigned int io_len = min_t(unsigned int, len, PAGE_SIZE);
> +
> + __bio_add_page(bio, page, io_len, 0);
> + len -= io_len;
> + }

I thought we were going to use the huge_zero_page for this?

2024-02-27 09:51:07

by Pankaj Raghav (Samsung)

Subject: Re: [PATCH 10/13] iomap: fix iomap_dio_zero() for fs bs > system page size

> I thought we were going to use the huge_zero_page for this?

Yes. We discussed that huge_zero_page might fail, so we concluded that
we needed an api that can return arbitrary folio order that will not
fail:
```
your point about it possibly failing is correct. so i think we need an
api which definitely returns a folio, but it might be of arbitrary
order.
```

I couldn't come up with implementing your latter suggestion, so I
informed darrick that let's use this patch for now, and add the
arbitrary folio order with zero as a later enhancement.

If we want to use mm_huge_zero_page, then this should work:

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index 04f6c5548136..b6a3f52f48da 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -237,10 +237,17 @@ static void iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
{
struct inode *inode = file_inode(dio->iocb->ki_filp);
struct page *page = ZERO_PAGE(0);
+ struct folio *folio = NULL;
struct bio *bio;

WARN_ON_ONCE(len > (BIO_MAX_VECS * PAGE_SIZE));

+ if (len > PAGE_SIZE) {
+ page = mm_get_huge_zero_page(current->mm);
+ if (!page)
+ page = ZERO_PAGE(0);
+ }
+
bio = iomap_dio_alloc_bio(iter, dio, BIO_MAX_VECS,
REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
@@ -249,13 +256,15 @@ static void iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
bio->bi_private = dio;
bio->bi_end_io = iomap_dio_bio_end_io;
+ folio = page_folio(page);

while (len) {
- unsigned int io_len = min_t(unsigned int, len, PAGE_SIZE);
+ size_t size = min(len, folio_size(folio));

- __bio_add_page(bio, page, io_len, 0);
- len -= io_len;
+ bio_add_folio_nofail(bio, folio, size, 0);
+ len -= size;
}
+
iomap_dio_submit_bio(iter, dio, bio, pos);
}

Let me know if we should go for this or let's keep the original patch
and add a ZERO_FOLIO_ORDER API that will not fail and use it as a later
enhancement.