| Message ID | 20240529134509.120826-8-kernel@pankajraghav.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | enable bs > ps in XFS | expand |
On Wed, May 29, 2024 at 03:45:05PM +0200, Pankaj Raghav (Samsung) wrote: > From: Pankaj Raghav <p.raghav@samsung.com> > > iomap_dio_zero() will pad a fs block with zeroes if the direct IO size > < fs block size. iomap_dio_zero() has an implicit assumption that fs block > size < page_size. This is true for most filesystems at the moment. > > If the block size > page size, this will send the contents of the page > next to zero page(as len > PAGE_SIZE) to the underlying block device, > causing FS corruption. > > iomap is a generic infrastructure and it should not make any assumptions > about the fs block size and the page size of the system. > > Signed-off-by: Pankaj Raghav <p.raghav@samsung.com> > --- > > After disucssing a bit in LSFMM about this, it was clear that using a > PMD sized zero folio might not be a good idea[0], especially in platforms > with 64k base page size, the huge zero folio can be as high as > 512M just for zeroing small block sizes in the direct IO path. > > The idea to use iomap_init to allocate 64k zero buffer was suggested by > Dave Chinner as it gives decent tradeoff between memory usage and efficiency. > > This is a good enough solution for now as moving beyond 64k block size > in XFS might take a while. We can work on a more generic solution in the > future to offer different sized zero folio that can go beyond 64k. > > [0] https://lore.kernel.org/linux-fsdevel/ZkdcAsENj2mBHh91@casper.infradead.org/ > > fs/internal.h | 8 ++++++++ > fs/iomap/buffered-io.c | 5 +++++ > fs/iomap/direct-io.c | 9 +++++++-- > 3 files changed, 20 insertions(+), 2 deletions(-) > > diff --git a/fs/internal.h b/fs/internal.h > index 84f371193f74..18eedbb82c50 100644 > --- a/fs/internal.h > +++ b/fs/internal.h > @@ -35,6 +35,14 @@ static inline void bdev_cache_init(void) > int __block_write_begin_int(struct folio *folio, loff_t pos, unsigned len, > get_block_t *get_block, const struct iomap *iomap); > > +/* > + * iomap/buffered-io.c > + */ > + > +#define ZERO_FSB_SIZE (65536) > +#define ZERO_FSB_ORDER (get_order(ZERO_FSB_SIZE)) > +extern struct page *zero_fs_block; This is really iomap direct IO private stuff. It should be visible anywhere else... > + > /* > * char_dev.c > */ > diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c > index c5802a459334..2c0149c827cd 100644 > --- a/fs/iomap/buffered-io.c > +++ b/fs/iomap/buffered-io.c > @@ -42,6 +42,7 @@ struct iomap_folio_state { > }; > > static struct bio_set iomap_ioend_bioset; > +struct page *zero_fs_block; > > static inline bool ifs_is_fully_uptodate(struct folio *folio, > struct iomap_folio_state *ifs) > @@ -1998,6 +1999,10 @@ EXPORT_SYMBOL_GPL(iomap_writepages); > > static int __init iomap_init(void) > { > + zero_fs_block = alloc_pages(GFP_KERNEL | __GFP_ZERO, ZERO_FSB_ORDER); > + if (!zero_fs_block) > + return -ENOMEM; > + > return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE), > offsetof(struct iomap_ioend, io_bio), > BIOSET_NEED_BVECS); just create an iomap_dio_init() function in iomap/direct-io.c and call that from here. Then everything can be private to iomap/direct-io.c... -Dave.
On 5/29/24 15:45, Pankaj Raghav (Samsung) wrote: > From: Pankaj Raghav <p.raghav@samsung.com> > > iomap_dio_zero() will pad a fs block with zeroes if the direct IO size > < fs block size. iomap_dio_zero() has an implicit assumption that fs block > size < page_size. This is true for most filesystems at the moment. > > If the block size > page size, this will send the contents of the page > next to zero page(as len > PAGE_SIZE) to the underlying block device, > causing FS corruption. > > iomap is a generic infrastructure and it should not make any assumptions > about the fs block size and the page size of the system. > > Signed-off-by: Pankaj Raghav <p.raghav@samsung.com> > --- > > After disucssing a bit in LSFMM about this, it was clear that using a > PMD sized zero folio might not be a good idea[0], especially in platforms > with 64k base page size, the huge zero folio can be as high as > 512M just for zeroing small block sizes in the direct IO path. > > The idea to use iomap_init to allocate 64k zero buffer was suggested by > Dave Chinner as it gives decent tradeoff between memory usage and efficiency. > > This is a good enough solution for now as moving beyond 64k block size > in XFS might take a while. We can work on a more generic solution in the > future to offer different sized zero folio that can go beyond 64k. > > [0] https://lore.kernel.org/linux-fsdevel/ZkdcAsENj2mBHh91@casper.infradead.org/ > Reviewed-by: Hannes Reinecke <hare@suse.de> Cheers, Hannes
> > static int __init iomap_init(void) > > { > > + zero_fs_block = alloc_pages(GFP_KERNEL | __GFP_ZERO, ZERO_FSB_ORDER); > > + if (!zero_fs_block) > > + return -ENOMEM; > > + > > return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE), > > offsetof(struct iomap_ioend, io_bio), > > BIOSET_NEED_BVECS); > > just create an iomap_dio_init() function in iomap/direct-io.c > and call that from here. Then everything can be private to > iomap/direct-io.c... Sounds good :)
diff --git a/fs/internal.h b/fs/internal.h index 84f371193f74..18eedbb82c50 100644 --- a/fs/internal.h +++ b/fs/internal.h @@ -35,6 +35,14 @@ static inline void bdev_cache_init(void) int __block_write_begin_int(struct folio *folio, loff_t pos, unsigned len, get_block_t *get_block, const struct iomap *iomap); +/* + * iomap/buffered-io.c + */ + +#define ZERO_FSB_SIZE (65536) +#define ZERO_FSB_ORDER (get_order(ZERO_FSB_SIZE)) +extern struct page *zero_fs_block; + /* * char_dev.c */ diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c index c5802a459334..2c0149c827cd 100644 --- a/fs/iomap/buffered-io.c +++ b/fs/iomap/buffered-io.c @@ -42,6 +42,7 @@ struct iomap_folio_state { }; static struct bio_set iomap_ioend_bioset; +struct page *zero_fs_block; static inline bool ifs_is_fully_uptodate(struct folio *folio, struct iomap_folio_state *ifs) @@ -1998,6 +1999,10 @@ EXPORT_SYMBOL_GPL(iomap_writepages); static int __init iomap_init(void) { + zero_fs_block = alloc_pages(GFP_KERNEL | __GFP_ZERO, ZERO_FSB_ORDER); + if (!zero_fs_block) + return -ENOMEM; + return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE), offsetof(struct iomap_ioend, io_bio), BIOSET_NEED_BVECS); diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c index f3b43d223a46..50c2bca8a347 100644 --- a/fs/iomap/direct-io.c +++ b/fs/iomap/direct-io.c @@ -236,17 +236,22 @@ static void iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio, loff_t pos, unsigned len) { struct inode *inode = file_inode(dio->iocb->ki_filp); - struct page *page = ZERO_PAGE(0); struct bio *bio; + /* + * Max block size supported is 64k + */ + WARN_ON_ONCE(len > ZERO_FSB_SIZE); + bio = iomap_dio_alloc_bio(iter, dio, 1, 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); + __bio_add_page(bio, zero_fs_block, len, 0); iomap_dio_submit_bio(iter, dio, bio, pos); }