Could we get an IOCB_NO_READ_HOLE?

Could we get an IOCB_NO_READ_HOLE?

[David Howells]

Hi Christoph,

Would it be possible to get an IOCB_NO_READ_HOLE flag that causes a read to
either fail entirely if there's a hole in the file or to stop at the hole,
possibly returning -ENODATA if the hole is at the front of the file?

Looking at iomap_dio_iter(), IOMAP_HOLE should be enabled in
iomap_iter::iomap.type for this?  Is it that simple?

David

Re: Could we get an IOCB_NO_READ_HOLE?

[David Howells]

David Howells <dhowells@redhat.com> wrote:

> Would it be possible to get an IOCB_NO_READ_HOLE flag that causes a read to
> either fail entirely if there's a hole in the file or to stop at the hole,
> possibly returning -ENODATA if the hole is at the front of the file?
> 
> Looking at iomap_dio_iter(), IOMAP_HOLE should be enabled in
> iomap_iter::iomap.type for this?  Is it that simple?

Actually, that's not the right thing.  How about the attached - at least for
direct I/O?

David
---
commit 522d2834f9994b82b1fa1f1eeeb48ede16b327c7
Author: David Howells <dhowells@redhat.com>
Date:   Fri Sep 17 11:33:41 2021 +0100

    iomap: Implement IOCB_NO_READ_HOLE

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index 4ecd255e0511..d2309dec27c4 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -18,6 +18,7 @@
  * Private flags for iomap_dio, must not overlap with the public ones in
  * iomap.h:
  */
+#define IOMAP_DIO_NO_READ_HOLE	(1 << 27)
 #define IOMAP_DIO_WRITE_FUA	(1 << 28)
 #define IOMAP_DIO_NEED_SYNC	(1 << 29)
 #define IOMAP_DIO_WRITE		(1 << 30)
@@ -412,6 +413,8 @@ static loff_t iomap_dio_iter(const struct iomap_iter *iter,
 	case IOMAP_HOLE:
 		if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
 			return -EIO;
+		if (dio->flags & IOMAP_DIO_NO_READ_HOLE)
+			return dio->size ? 0 : -ENODATA;
 		return iomap_dio_hole_iter(iter, dio);
 	case IOMAP_UNWRITTEN:
 		if (!(dio->flags & IOMAP_DIO_WRITE))
@@ -503,6 +506,8 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 
 		if (iter_is_iovec(iter))
 			dio->flags |= IOMAP_DIO_DIRTY;
+		if (iocb->ki_flags & IOCB_NO_READ_HOLE)
+			dio->flags |= IOMAP_DIO_NO_READ_HOLE;
 	} else {
 		iomi.flags |= IOMAP_WRITE;
 		dio->flags |= IOMAP_DIO_WRITE;
diff --git a/include/linux/fs.h b/include/linux/fs.h
index e7a633353fd2..f4c8ca22531d 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -321,6 +321,7 @@ enum rw_hint {
 #define IOCB_NOIO		(1 << 20)
 /* can use bio alloc cache */
 #define IOCB_ALLOC_CACHE	(1 << 21)
+#define IOCB_NO_READ_HOLE	(1 << 22)	/* Don't read from a hole */
 
 struct kiocb {
 	struct file		*ki_filp;

Re: Could we get an IOCB_NO_READ_HOLE?

[Christoph Hellwig]

On Fri, Sep 17, 2021 at 11:43:59AM +0100, David Howells wrote:
> David Howells <dhowells@redhat.com> wrote:
> 
> > Would it be possible to get an IOCB_NO_READ_HOLE flag that causes a read to
> > either fail entirely if there's a hole in the file or to stop at the hole,
> > possibly returning -ENODATA if the hole is at the front of the file?
> > 
> > Looking at iomap_dio_iter(), IOMAP_HOLE should be enabled in
> > iomap_iter::iomap.type for this?  Is it that simple?
> 
> Actually, that's not the right thing.  How about the attached - at least for
> direct I/O?

This looks pretty reasonable.  We'll just need to make sure to reject
the flag for the many file operations instances that do not support it.