ext4: Add atomic write support for DIO

This patch adds base support for atomic writes via statx getattr. On bs < ps systems, we can create FS with say bs of 16k. That means both atomic write min and max unit can be set to 16k for supporting atomic writes.

Later patches adds support for bigalloc as well so that ext4 can also support doing atomic writes for bs = ps systems.

Co-developed-by Ojaswin Mujoo <[email protected]>

Signed-off-by Ojaswin Mujoo <[email protected]>

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
  fs/ext4/ext4.h  |  7 ++++++-
  fs/ext4/inode.c | 14 ++++++++++++++
  fs/ext4/super.c | 32 ++++++++++++++++++++++++++++++++
  3 files changed, 52 insertions(+), 1 deletion(-)

fs/ext4/ext4.h +4 -0

index 44b0d418143c..a41e56c2c628 100644
@@ -1729,6 +1729,10 @@ struct ext4_sb_info {
  	 */
  	struct work_struct s_sb_upd_work;
 
+	/* Atomic write unit values */
+	unsigned int fs_awu_min;
+	unsigned int fs_awu_max;
+
  	/* Ext4 fast commit sub transaction ID */
  	atomic_t s_fc_subtid;
 

fs/ext4/ext4.h +0 -0

  };
 
  static inline void ext4_set_mount_flag(struct super_block *sb, int bit)

fs/ext4/inode.c +0 -0

index 54bdd4884fe6..897c028d5bc9 100644

fs/ext4/inode.c +9 -0

+			awu_min = sbi->fs_awu_min;
+			awu_max = sbi->fs_awu_max;
+		} else {
+			awu_min = awu_max = 0;
+		}
+
+		generic_fill_statx_atomic_writes(stat, awu_min, awu_max);
+	}
+
  	flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
  	if (flags & EXT4_APPEND_FL)
  		stat->attributes |= STATX_ATTR_APPEND;

fs/ext4/super.c +1 -0

index 16a4ce704460..f5c075aff060 100644
@@ -4425,6 +4425,37 @@ static int ext4_handle_clustersize(struct super_block *sb)
  	return 0;
  }
 
+/*

Quoted 32 lines · Ritesh Harjani (IBM)

This patch adds base support for atomic writes via statx getattr. On bs < ps systems, we can create FS with say bs of 16k. That means both atomic write min and max unit can be set to 16k for supporting atomic writes.

Later patches adds support for bigalloc as well so that ext4 can also support doing atomic writes for bs = ps systems.

Co-developed-by Ojaswin Mujoo <[email protected]>

Signed-off-by Ojaswin Mujoo <[email protected]>

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
  fs/ext4/ext4.h  |  7 ++++++-
  fs/ext4/inode.c | 14 ++++++++++++++
  fs/ext4/super.c | 32 ++++++++++++++++++++++++++++++++
  3 files changed, 52 insertions(+), 1 deletion(-)

fs/ext4/ext4.h +4 -0

index 44b0d418143c..a41e56c2c628 100644
@@ -1729,6 +1729,10 @@ struct ext4_sb_info {
  	 */
  	struct work_struct s_sb_upd_work;
 
+	/* Atomic write unit values */
+	unsigned int fs_awu_min;
+	unsigned int fs_awu_max;
+
  	/* Ext4 fast commit sub transaction ID */
  	atomic_t s_fc_subtid;
 

fs/ext4/ext4.h static inline int ext4_valid_inum(struct super_block *sb, unsigned long ino) · line 1828

Quoted 7 lines

fs/ext4/ext4.h +2 -1

@@ -1820,7 +1824,8 @@ static inline int ext4_valid_inum(struct super_block *sb, unsigned long ino)
   */
  enum {
  	EXT4_MF_MNTDIR_SAMPLED,
-	EXT4_MF_FC_INELIGIBLE	/* Fast commit ineligible */
+	EXT4_MF_FC_INELIGIBLE,	/* Fast commit ineligible */
+	EXT4_MF_ATOMIC_WRITE	/* Supports atomic write */

Does this flag really buy us much?

Quoted 11 lines

fs/ext4/ext4.h +0 -0

  };
 
  static inline void ext4_set_mount_flag(struct super_block *sb, int bit)

fs/ext4/inode.c +0 -0

index 54bdd4884fe6..897c028d5bc9 100644
@@ -5578,6 +5578,20 @@ int ext4_getattr(struct mnt_idmap *idmap, const struct path *path,
  		}
  	}
 

fs/ext4/inode.c int ext4_getattr(struct mnt_idmap *idmap, const struct path *path, · line 5585

Quoted 5 lines

fs/ext4/inode.c +5 -0

+	if (S_ISREG(inode->i_mode) && (request_mask & STATX_WRITE_ATOMIC)) {
+		struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
+		unsigned int awu_min, awu_max;
+
+		if (ext4_test_mount_flag(inode->i_sb, EXT4_MF_ATOMIC_WRITE)) {

I'd use ext4_inode_can_atomicwrite() here, similar to what is done for xfs

Quoted 21 lines

fs/ext4/inode.c +9 -0

+			awu_min = sbi->fs_awu_min;
+			awu_max = sbi->fs_awu_max;
+		} else {
+			awu_min = awu_max = 0;
+		}
+
+		generic_fill_statx_atomic_writes(stat, awu_min, awu_max);
+	}
+
  	flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
  	if (flags & EXT4_APPEND_FL)
  		stat->attributes |= STATX_ATTR_APPEND;

fs/ext4/super.c +1 -0

index 16a4ce704460..f5c075aff060 100644
@@ -4425,6 +4425,37 @@ static int ext4_handle_clustersize(struct super_block *sb)
  	return 0;
  }
 
+/*

Quoted 33 lines · John Garry

Quoted 32 lines · Ritesh Harjani (IBM)

This patch adds base support for atomic writes via statx getattr. On bs < ps systems, we can create FS with say bs of 16k. That means both atomic write min and max unit can be set to 16k for supporting atomic writes.

Later patches adds support for bigalloc as well so that ext4 can also support doing atomic writes for bs = ps systems.

Co-developed-by Ojaswin Mujoo <[email protected]>

Signed-off-by Ojaswin Mujoo <[email protected]>

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
  fs/ext4/ext4.h  |  7 ++++++-
  fs/ext4/inode.c | 14 ++++++++++++++
  fs/ext4/super.c | 32 ++++++++++++++++++++++++++++++++
  3 files changed, 52 insertions(+), 1 deletion(-)

fs/ext4/ext4.h +4 -0

index 44b0d418143c..a41e56c2c628 100644
@@ -1729,6 +1729,10 @@ struct ext4_sb_info {
  	 */
  	struct work_struct s_sb_upd_work;
 
+	/* Atomic write unit values */
+	unsigned int fs_awu_min;
+	unsigned int fs_awu_max;
+
  	/* Ext4 fast commit sub transaction ID */
  	atomic_t s_fc_subtid;
 

Quoted 10 lines

fs/ext4/ext4.h static inline int ext4_valid_inum(struct super_block *sb, unsigned long ino) · line 1828

Quoted 7 lines

fs/ext4/ext4.h +2 -1

@@ -1820,7 +1824,8 @@ static inline int ext4_valid_inum(struct super_block *sb, unsigned long ino)
   */
  enum {
  	EXT4_MF_MNTDIR_SAMPLED,
-	EXT4_MF_FC_INELIGIBLE	/* Fast commit ineligible */
+	EXT4_MF_FC_INELIGIBLE,	/* Fast commit ineligible */
+	EXT4_MF_ATOMIC_WRITE	/* Supports atomic write */

Does this flag really buy us much?

I felt it is cleaner this way than comparing non-zero values of fs_awu_min and fs_awu_max.

Now that you pointed at it - Maybe a question for others who might have the history of which one to use when - or do we think there is a scope of merging the two into just one as a later cleanup?

I know that s_mount_flags was added for fastcommit and it needed the state manipulations to be done in atomic way. Similarly s_ext4_flags also was renamed from s_resize_flags for more general purpose use. Both of these looks like could be merged isn't it?

Quoted 11 lines

Quoted 11 lines

fs/ext4/ext4.h +0 -0

  };
 
  static inline void ext4_set_mount_flag(struct super_block *sb, int bit)

fs/ext4/inode.c +0 -0

index 54bdd4884fe6..897c028d5bc9 100644
@@ -5578,6 +5578,20 @@ int ext4_getattr(struct mnt_idmap *idmap, const struct path *path,
  		}
  	}
 

Quoted 8 lines

fs/ext4/inode.c int ext4_getattr(struct mnt_idmap *idmap, const struct path *path, · line 5585

Quoted 5 lines

fs/ext4/inode.c +5 -0

+	if (S_ISREG(inode->i_mode) && (request_mask & STATX_WRITE_ATOMIC)) {
+		struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
+		unsigned int awu_min, awu_max;
+
+		if (ext4_test_mount_flag(inode->i_sb, EXT4_MF_ATOMIC_WRITE)) {

I'd use ext4_inode_can_atomicwrite() here, similar to what is done for xfs

Sure since it is inode operation, we can check against ext4_inode_can_atomicwrite().

Quoted 21 lines

Quoted 21 lines

fs/ext4/inode.c +9 -0

+			awu_min = sbi->fs_awu_min;
+			awu_max = sbi->fs_awu_max;
+		} else {
+			awu_min = awu_max = 0;
+		}
+
+		generic_fill_statx_atomic_writes(stat, awu_min, awu_max);
+	}
+
  	flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
  	if (flags & EXT4_APPEND_FL)
  		stat->attributes |= STATX_ATTR_APPEND;

fs/ext4/super.c +1 -0

index 16a4ce704460..f5c075aff060 100644
@@ -4425,6 +4425,37 @@ static int ext4_handle_clustersize(struct super_block *sb)
  	return 0;
  }
 
+/*

Quoted 46 lines · Ritesh Harjani

Quoted 43 lines · John Garry

Quoted 33 lines · Ritesh Harjani (IBM)

This patch adds base support for atomic writes via statx getattr. On bs < ps systems, we can create FS with say bs of 16k. That means both atomic write min and max unit can be set to 16k for supporting atomic writes.

Later patches adds support for bigalloc as well so that ext4 can also support doing atomic writes for bs = ps systems.

Co-developed-by Ojaswin Mujoo <[email protected]>

Signed-off-by Ojaswin Mujoo <[email protected]>

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
  fs/ext4/ext4.h  |  7 ++++++-
  fs/ext4/inode.c | 14 ++++++++++++++
  fs/ext4/super.c | 32 ++++++++++++++++++++++++++++++++
  3 files changed, 52 insertions(+), 1 deletion(-)

fs/ext4/ext4.h +4 -0

index 44b0d418143c..a41e56c2c628 100644
@@ -1729,6 +1729,10 @@ struct ext4_sb_info {
  	 */
  	struct work_struct s_sb_upd_work;
 
+	/* Atomic write unit values */
+	unsigned int fs_awu_min;
+	unsigned int fs_awu_max;
+
  	/* Ext4 fast commit sub transaction ID */
  	atomic_t s_fc_subtid;
 

fs/ext4/ext4.h static inline int ext4_valid_inum(struct super_block *sb, unsigned long ino) · line 1828

Quoted 6 lines

fs/ext4/ext4.h +2 -1

@@ -1820,7 +1824,8 @@ static inline int ext4_valid_inum(struct super_block *sb, unsigned long ino)
   */
  enum {
  	EXT4_MF_MNTDIR_SAMPLED,
-	EXT4_MF_FC_INELIGIBLE	/* Fast commit ineligible */
+	EXT4_MF_FC_INELIGIBLE,	/* Fast commit ineligible */
+	EXT4_MF_ATOMIC_WRITE	/* Supports atomic write */

Does this flag really buy us much?

Quoted 2 lines

I felt it is cleaner this way than comparing non-zero values of fs_awu_min and fs_awu_max.

What does it mean when MF_ATOMIC_WRITE is set and fs_awu_* are zero? The awu values don't change at runtime, so I think you can save yourself an atomic test by checking (non-atomically) for awu_min>0.

(I don't know anything about the flags, those came after my time iirc.)

--D

Quoted 56 lines

Now that you pointed at it - Maybe a question for others who might have the history of which one to use when - or do we think there is a scope of merging the two into just one as a later cleanup?

I know that s_mount_flags was added for fastcommit and it needed the state manipulations to be done in atomic way. Similarly s_ext4_flags also was renamed from s_resize_flags for more general purpose use. Both of these looks like could be merged isn't it?

Quoted 11 lines

Quoted 11 lines

fs/ext4/ext4.h +0 -0

  };
 
  static inline void ext4_set_mount_flag(struct super_block *sb, int bit)

fs/ext4/inode.c +0 -0

index 54bdd4884fe6..897c028d5bc9 100644
@@ -5578,6 +5578,20 @@ int ext4_getattr(struct mnt_idmap *idmap, const struct path *path,
  		}
  	}
 

Quoted 8 lines

fs/ext4/inode.c int ext4_getattr(struct mnt_idmap *idmap, const struct path *path, · line 5585

Quoted 5 lines

fs/ext4/inode.c +5 -0

+	if (S_ISREG(inode->i_mode) && (request_mask & STATX_WRITE_ATOMIC)) {
+		struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
+		unsigned int awu_min, awu_max;
+
+		if (ext4_test_mount_flag(inode->i_sb, EXT4_MF_ATOMIC_WRITE)) {

I'd use ext4_inode_can_atomicwrite() here, similar to what is done for xfs

Sure since it is inode operation, we can check against ext4_inode_can_atomicwrite().

Quoted 21 lines

Quoted 21 lines

fs/ext4/inode.c +9 -0

+			awu_min = sbi->fs_awu_min;
+			awu_max = sbi->fs_awu_max;
+		} else {
+			awu_min = awu_max = 0;
+		}
+
+		generic_fill_statx_atomic_writes(stat, awu_min, awu_max);
+	}
+
  	flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
  	if (flags & EXT4_APPEND_FL)
  		stat->attributes |= STATX_ATTR_APPEND;

fs/ext4/super.c +1 -0

index 16a4ce704460..f5c075aff060 100644
@@ -4425,6 +4425,37 @@ static int ext4_handle_clustersize(struct super_block *sb)
  	return 0;
  }
 
+/*

Let's validate using generic_atomic_write_valid() in ext4_file_write_iter() if the write request has IOCB_ATOMIC set.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
  fs/ext4/file.c | 14 ++++++++++++++
  1 file changed, 14 insertions(+)

fs/ext4/file.c +0 -0

index f14aed14b9cf..b06c5d34bbd2 100644
@@ -692,6 +692,20 @@ ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
  	if (IS_DAX(inode))

fs/ext4/file.c +6 -0

+
+		ret = generic_atomic_write_valid(iocb, from);
+		if (ret)
+			return ret;
+	}
+
  	if (iocb->ki_flags & IOCB_DIRECT)
  		return ext4_dio_write_iter(iocb, from);
  	else

Quoted 19 lines · Ritesh Harjani (IBM)

Let's validate using generic_atomic_write_valid() in ext4_file_write_iter() if the write request has IOCB_ATOMIC set.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
  fs/ext4/file.c | 14 ++++++++++++++
  1 file changed, 14 insertions(+)

fs/ext4/file.c +3 -0

index f14aed14b9cf..b06c5d34bbd2 100644
@@ -692,6 +692,20 @@ ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
  	if (IS_DAX(inode))
  		return ext4_dax_write_iter(iocb, from);
  #endif
+
+	if (iocb->ki_flags & IOCB_ATOMIC) {
+		size_t len = iov_iter_count(from);

fs/ext4/file.c ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 702

Quoted 5 lines

fs/ext4/file.c +5 -0

+		int ret;
+
+		if (!IS_ALIGNED(len, EXT4_SB(inode->i_sb)->fs_awu_min) ||
+			len > EXT4_SB(inode->i_sb)->fs_awu_max)
+			return -EINVAL;

this looks ok, but the IS_ALIGNED() check looks odd. I am not sure why you don't just check that fs_awu_max >= len >= fs_awu_min

fs/ext4/file.c ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 711

Quoted 9 lines

fs/ext4/file.c +6 -0

+
+		ret = generic_atomic_write_valid(iocb, from);
+		if (ret)
+			return ret;
+	}
+
  	if (iocb->ki_flags & IOCB_DIRECT)
  		return ext4_dio_write_iter(iocb, from);
  	else

Quoted 20 lines · John Garry

Quoted 19 lines · Ritesh Harjani (IBM)

Let's validate using generic_atomic_write_valid() in ext4_file_write_iter() if the write request has IOCB_ATOMIC set.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
  fs/ext4/file.c | 14 ++++++++++++++
  1 file changed, 14 insertions(+)

fs/ext4/file.c +3 -0

index f14aed14b9cf..b06c5d34bbd2 100644
@@ -692,6 +692,20 @@ ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
  	if (IS_DAX(inode))
  		return ext4_dax_write_iter(iocb, from);
  #endif
+
+	if (iocb->ki_flags & IOCB_ATOMIC) {
+		size_t len = iov_iter_count(from);

Quoted 9 lines

fs/ext4/file.c ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 702

Quoted 5 lines

fs/ext4/file.c +5 -0

+		int ret;
+
+		if (!IS_ALIGNED(len, EXT4_SB(inode->i_sb)->fs_awu_min) ||
+			len > EXT4_SB(inode->i_sb)->fs_awu_max)
+			return -EINVAL;

this looks ok, but the IS_ALIGNED() check looks odd. I am not sure why you don't just check that fs_awu_max >= len >= fs_awu_min

I guess this was just a stricter check. But we anyways have power_of_2 and other checks in generic_atomic_write_valid(). So it does not matter.

I can change this in v2.

Thanks!

Quoted 9 lines

fs/ext4/file.c ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 711

Quoted 9 lines

fs/ext4/file.c +6 -0

+
+		ret = generic_atomic_write_valid(iocb, from);
+		if (ret)
+			return ret;
+	}
+
  	if (iocb->ki_flags & IOCB_DIRECT)
  		return ext4_dio_write_iter(iocb, from);
  	else

-ritesh

Quoted 32 lines · Ritesh Harjani

Quoted 20 lines · John Garry

Quoted 19 lines · Ritesh Harjani (IBM)

Let's validate using generic_atomic_write_valid() in ext4_file_write_iter() if the write request has IOCB_ATOMIC set.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
  fs/ext4/file.c | 14 ++++++++++++++
  1 file changed, 14 insertions(+)

fs/ext4/file.c +3 -0

index f14aed14b9cf..b06c5d34bbd2 100644
@@ -692,6 +692,20 @@ ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
  	if (IS_DAX(inode))
  		return ext4_dax_write_iter(iocb, from);
  #endif
+
+	if (iocb->ki_flags & IOCB_ATOMIC) {
+		size_t len = iov_iter_count(from);

Quoted 9 lines

fs/ext4/file.c ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 702

Quoted 5 lines

fs/ext4/file.c +5 -0

+		int ret;
+
+		if (!IS_ALIGNED(len, EXT4_SB(inode->i_sb)->fs_awu_min) ||
+			len > EXT4_SB(inode->i_sb)->fs_awu_max)
+			return -EINVAL;

this looks ok, but the IS_ALIGNED() check looks odd. I am not sure why you don't just check that fs_awu_max >= len >= fs_awu_min

Quoted 4 lines

I guess this was just a stricter check. But we anyways have power_of_2 and other checks in generic_atomic_write_valid(). So it does not matter.

I can change this in v2.

Also please fix the weird indenting in the if test:

		if (len < EXT4_SB(inode->i_sb)->fs_awu_min) ||
		    len > EXT4_SB(inode->i_sb)->fs_awu_max)
			return -EINVAL;

--D

Quoted 15 lines

Thanks!

Quoted 9 lines

fs/ext4/file.c ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 711

Quoted 9 lines

fs/ext4/file.c +6 -0

+
+		ret = generic_atomic_write_valid(iocb, from);
+		if (ret)
+			return ret;
+	}
+
  	if (iocb->ki_flags & IOCB_DIRECT)
  		return ext4_dio_write_iter(iocb, from);
  	else

-ritesh

iomap will not return -ENOTBLK in case of dio atomic writes. But let's also add a WARN_ON_ONCE and return -EIO as a safety net.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
 fs/ext4/file.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

fs/ext4/file.c +0 -0

index f9516121a036..af6ebd0ac0d6 100644

fs/ext4/file.c +5 -0

+		 */
+		if (WARN_ON_ONCE(iocb->ki_flags & IOCB_ATOMIC))
+			ret = -EIO;
+	}
+
 	if (extend) {
 		/*
 		 * We always perform extending DIO write synchronously so by

Signature · 1 line

2.46.0

Quoted 17 lines · Ritesh Harjani (IBM)

iomap will not return -ENOTBLK in case of dio atomic writes. But let's also add a WARN_ON_ONCE and return -EIO as a safety net.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
 fs/ext4/file.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

fs/ext4/file.c +0 -1

index f9516121a036..af6ebd0ac0d6 100644
@@ -576,8 +576,16 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 		iomap_ops = &ext4_iomap_overwrite_ops;
 	ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 			   dio_flags, NULL, 0);
-	if (ret == -ENOTBLK)

fs/ext4/file.c static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 583

Quoted 5 lines

fs/ext4/file.c +4 -0

+	if (ret == -ENOTBLK) {
 		ret = 0;
+		/*
+		 * iomap will never return -ENOTBLK if write fails for atomic
+		 * write. But let's just add a safety net.

I think it can if the pagecache invalidation fails, so you really do need the safety net.

I suspect that the xfs version of this series
needs it too, though it may have fallen out?

Reviewed-by Darrick J. Wong <[email protected]>

--D

Quoted 12 lines

fs/ext4/file.c +5 -0

+		 */
+		if (WARN_ON_ONCE(iocb->ki_flags & IOCB_ATOMIC))
+			ret = -EIO;
+	}
+
 	if (extend) {
 		/*
 		 * We always perform extending DIO write synchronously so by

Signature · 1 line

2.46.0

iomap will not return -ENOTBLK in case of dio atomic writes. But let's also add a WARN_ON_ONCE and return -EIO as a safety net.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
 fs/ext4/file.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

fs/ext4/file.c +0 -0

index f9516121a036..af6ebd0ac0d6 100644
@@ -576,8 +576,16 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 		iomap_ops = &ext4_iomap_overwrite_ops;
 	ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 			   dio_flags, NULL, 0);

Quoted 21 lines · Ritesh Harjani (IBM)

iomap will not return -ENOTBLK in case of dio atomic writes. But let's also add a WARN_ON_ONCE and return -EIO as a safety net.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
 fs/ext4/file.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

fs/ext4/file.c +3 -1

index f9516121a036..af6ebd0ac0d6 100644
@@ -576,8 +576,16 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 		iomap_ops = &ext4_iomap_overwrite_ops;
 	ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 			   dio_flags, NULL, 0);
-	if (ret == -ENOTBLK)
+	if (ret == -ENOTBLK) {
 		ret = 0;
+		/*
+		 * iomap will never return -ENOTBLK if write fails for atomic

fs/ext4/file.c static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 587

Quoted 5 lines

fs/ext4/file.c +5 -0

+		 * write. But let's just add a safety net.
+		 */
+		if (WARN_ON_ONCE(iocb->ki_flags & IOCB_ATOMIC))
+			ret = -EIO;
+	}

Why can't the iomap code return EIO in this case for IOCB_ATOMIC? That way we don't have to put this logic into every filesystem.

When/if we start supporting atomic writes for buffered IO, then it's worth pushing this out to filesystems, but right now it doesn't seem necessary...

-Dave.

Signature · 2 lines

Dave Chinner
[email protected]

Hi Dave,

Quoted 22 lines · Dave Chinner

Quoted 21 lines · Ritesh Harjani (IBM)

iomap will not return -ENOTBLK in case of dio atomic writes. But let's also add a WARN_ON_ONCE and return -EIO as a safety net.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
 fs/ext4/file.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

fs/ext4/file.c +3 -1

index f9516121a036..af6ebd0ac0d6 100644
@@ -576,8 +576,16 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 		iomap_ops = &ext4_iomap_overwrite_ops;
 	ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 			   dio_flags, NULL, 0);
-	if (ret == -ENOTBLK)
+	if (ret == -ENOTBLK) {
 		ret = 0;
+		/*
+		 * iomap will never return -ENOTBLK if write fails for atomic

Quoted 8 lines

fs/ext4/file.c static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 587

Quoted 5 lines

fs/ext4/file.c +5 -0

+		 * write. But let's just add a safety net.
+		 */
+		if (WARN_ON_ONCE(iocb->ki_flags & IOCB_ATOMIC))
+			ret = -EIO;
+	}

Why can't the iomap code return EIO in this case for IOCB_ATOMIC? That way we don't have to put this logic into every filesystem.

This was origially intended as a safety net hence the WARN_ON_ONCE. Later Darrick pointed out that we still might have an unconverted condition in iomap which can return ENOTBLK for DIO atomic writes (page cache invalidation).

You pointed it right that it should be fixed in iomap. However do you think filesystems can still keep this as safety net (maybe no need of WARN_ON_ONCE).

Quoted 36 lines · Ritesh Harjani

Hi Dave,

Quoted 22 lines · Dave Chinner

Quoted 21 lines · Ritesh Harjani (IBM)

iomap will not return -ENOTBLK in case of dio atomic writes. But let's also add a WARN_ON_ONCE and return -EIO as a safety net.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
 fs/ext4/file.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

fs/ext4/file.c +3 -1

index f9516121a036..af6ebd0ac0d6 100644
@@ -576,8 +576,16 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 		iomap_ops = &ext4_iomap_overwrite_ops;
 	ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 			   dio_flags, NULL, 0);
-	if (ret == -ENOTBLK)
+	if (ret == -ENOTBLK) {
 		ret = 0;
+		/*
+		 * iomap will never return -ENOTBLK if write fails for atomic

Quoted 8 lines

fs/ext4/file.c static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 587

Quoted 5 lines

fs/ext4/file.c +5 -0

+		 * write. But let's just add a safety net.
+		 */
+		if (WARN_ON_ONCE(iocb->ki_flags & IOCB_ATOMIC))
+			ret = -EIO;
+	}

Why can't the iomap code return EIO in this case for IOCB_ATOMIC? That way we don't have to put this logic into every filesystem.

Quoted 4 lines

This was origially intended as a safety net hence the WARN_ON_ONCE. Later Darrick pointed out that we still might have an unconverted condition in iomap which can return ENOTBLK for DIO atomic writes (page cache invalidation).

Yes. That's my point - iomap knows that it's an atomic write, it knows that invalidation failed, and it knows that there is no such thing as buffered atomic writes. So there is no possible fallback here, and it should be returning EIO in the page cache invalidation failure case and not ENOTBLK.

Quoted 3 lines

You pointed it right that it should be fixed in iomap. However do you think filesystems can still keep this as safety net (maybe no need of WARN_ON_ONCE).

I don't see any point in adding "impossible to hit" checks into filesystems just in case some core infrastructure has a bug introduced....

-Dave.

Signature · 2 lines

Dave Chinner
[email protected]

Quoted 36 lines · Ritesh Harjani

Hi Dave,

Quoted 22 lines · Dave Chinner

Quoted 21 lines · Ritesh Harjani (IBM)

iomap will not return -ENOTBLK in case of dio atomic writes. But let's also add a WARN_ON_ONCE and return -EIO as a safety net.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
 fs/ext4/file.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

fs/ext4/file.c +3 -1

index f9516121a036..af6ebd0ac0d6 100644
@@ -576,8 +576,16 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 		iomap_ops = &ext4_iomap_overwrite_ops;
 	ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 			   dio_flags, NULL, 0);
-	if (ret == -ENOTBLK)
+	if (ret == -ENOTBLK) {
 		ret = 0;
+		/*
+		 * iomap will never return -ENOTBLK if write fails for atomic

Quoted 8 lines

fs/ext4/file.c static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 587

Quoted 5 lines

fs/ext4/file.c +5 -0

+		 * write. But let's just add a safety net.
+		 */
+		if (WARN_ON_ONCE(iocb->ki_flags & IOCB_ATOMIC))
+			ret = -EIO;
+	}

Why can't the iomap code return EIO in this case for IOCB_ATOMIC? That way we don't have to put this logic into every filesystem.

Quoted 4 lines

This was origially intended as a safety net hence the WARN_ON_ONCE. Later Darrick pointed out that we still might have an unconverted condition in iomap which can return ENOTBLK for DIO atomic writes (page cache invalidation).

Yes. That's my point - iomap knows that it's an atomic write, it knows that invalidation failed, and it knows that there is no such thing as buffered atomic writes. So there is no possible fallback here, and it should be returning EIO in the page cache invalidation failure case and not ENOTBLK.

Quoted 3 lines

You pointed it right that it should be fixed in iomap. However do you think filesystems can still keep this as safety net (maybe no need of WARN_ON_ONCE).

I don't see any point in adding "impossible to hit" checks into filesystems just in case some core infrastructure has a bug introduced....

Hi Dave,

Quoted 42 lines · Dave Chinner

Quoted 36 lines · Ritesh Harjani

Hi Dave,

Quoted 22 lines · Dave Chinner

Quoted 21 lines · Ritesh Harjani (IBM)

iomap will not return -ENOTBLK in case of dio atomic writes. But let's also add a WARN_ON_ONCE and return -EIO as a safety net.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
 fs/ext4/file.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

fs/ext4/file.c +3 -1

index f9516121a036..af6ebd0ac0d6 100644
@@ -576,8 +576,16 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 		iomap_ops = &ext4_iomap_overwrite_ops;
 	ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 			   dio_flags, NULL, 0);
-	if (ret == -ENOTBLK)
+	if (ret == -ENOTBLK) {
 		ret = 0;
+		/*
+		 * iomap will never return -ENOTBLK if write fails for atomic

Quoted 8 lines

fs/ext4/file.c static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 587

Quoted 5 lines

fs/ext4/file.c +5 -0

+		 * write. But let's just add a safety net.
+		 */
+		if (WARN_ON_ONCE(iocb->ki_flags & IOCB_ATOMIC))
+			ret = -EIO;
+	}

Why can't the iomap code return EIO in this case for IOCB_ATOMIC? That way we don't have to put this logic into every filesystem.

Quoted 4 lines

This was origially intended as a safety net hence the WARN_ON_ONCE. Later Darrick pointed out that we still might have an unconverted condition in iomap which can return ENOTBLK for DIO atomic writes (page cache invalidation).

Quoted 6 lines

Yes. That's my point - iomap knows that it's an atomic write, it knows that invalidation failed, and it knows that there is no such thing as buffered atomic writes. So there is no possible fallback here, and it should be returning EIO in the page cache invalidation failure case and not ENOTBLK.

So the iomap DIO can return following as return values which can make some filesystems fallback to buffered-io (if they implement fallback logic) -
(1) -ENOTBLK -> this is only returned for pagecache invalidation failure.
(2) 0 or partial write size -> This can never happen for atomic writes (since we are only allowing for single fsblock as of now).

Now looking at XFS, it never fallsback to buffered-io ever except just 2 cases -
1. When pagecache invalidation fails in iomap (can never happen for atomic writes)

2. On unaligned DIO writes to reflinked CoW (not possible for atomic writes)

So it anyways should never happen that XFS ever fallback to buffered-io for DIO atomic writes. Even today it does not fallback to buffered-io for non-atomic short DIO writes.

Quoted 4 lines

Quoted 3 lines

You pointed it right that it should be fixed in iomap. However do you think filesystems can still keep this as safety net (maybe no need of WARN_ON_ONCE).

Quoted 3 lines

I don't see any point in adding "impossible to hit" checks into filesystems just in case some core infrastructure has a bug introduced....

Yes, that is true for XFS. EXT4 however can return -ENOTBLK for short writes, though it should not happen for current atomic write case where we are only allowing for 1 fsblock.

I would still like to go with a WARN_ON_ONCE where we are calling ext4 buffered-io handling for DIO fallback writes. This is to catch any bugs even in future when we move to multi-fsblock case (until we have atomic write support for buffered-io).

Quoted 54 lines · Ritesh Harjani

Hi Dave,

Quoted 42 lines · Dave Chinner

Quoted 36 lines · Ritesh Harjani

Hi Dave,

Quoted 22 lines · Dave Chinner

Quoted 21 lines · Ritesh Harjani (IBM)

iomap will not return -ENOTBLK in case of dio atomic writes. But let's also add a WARN_ON_ONCE and return -EIO as a safety net.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
 fs/ext4/file.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

fs/ext4/file.c +3 -1

index f9516121a036..af6ebd0ac0d6 100644
@@ -576,8 +576,16 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 		iomap_ops = &ext4_iomap_overwrite_ops;
 	ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 			   dio_flags, NULL, 0);
-	if (ret == -ENOTBLK)
+	if (ret == -ENOTBLK) {
 		ret = 0;
+		/*
+		 * iomap will never return -ENOTBLK if write fails for atomic

Quoted 8 lines

fs/ext4/file.c static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from) · line 587

Quoted 5 lines

fs/ext4/file.c +5 -0

+		 * write. But let's just add a safety net.
+		 */
+		if (WARN_ON_ONCE(iocb->ki_flags & IOCB_ATOMIC))
+			ret = -EIO;
+	}

Why can't the iomap code return EIO in this case for IOCB_ATOMIC? That way we don't have to put this logic into every filesystem.

Quoted 4 lines

This was origially intended as a safety net hence the WARN_ON_ONCE. Later Darrick pointed out that we still might have an unconverted condition in iomap which can return ENOTBLK for DIO atomic writes (page cache invalidation).

Quoted 6 lines

Yes. That's my point - iomap knows that it's an atomic write, it knows that invalidation failed, and it knows that there is no such thing as buffered atomic writes. So there is no possible fallback here, and it should be returning EIO in the page cache invalidation failure case and not ENOTBLK.

Quoted 6 lines

So the iomap DIO can return following as return values which can make some filesystems fallback to buffered-io (if they implement fallback logic) -
(1) -ENOTBLK -> this is only returned for pagecache invalidation failure.
(2) 0 or partial write size -> This can never happen for atomic writes (since we are only allowing for single fsblock as of now).

Even when we allow multi-FSB atomic writes, the definition of atomic write is still "all or nothing". There is no scope for "short writes" when IOCB_ATOMIC is set - any condition that means we can't write the entire IO as a single bio, we need to abort and return EINVAL.

Hence -ENOTBLK should never be returned by iomap for atomic DIO writes - we need to say -EINVAL if the write could not be issued atomically for whatever reason it may be so the application knows that atomic IO submission was not possible for that IO.

Quoted 4 lines

Now looking at XFS, it never fallsback to buffered-io ever except just 2 cases -
1. When pagecache invalidation fails in iomap (can never happen for atomic writes)

Why can't this happen for atomic DIO writes? It's the same failure cases as for normal DIO writes, isn't it? (i.e. race with mmap writes)

My point is that if it's an atomic write, this failure should get turned into -EINVAL by the iomap code. We do not want a fallback to buffered IO when this situation happens for atomic IO.

Quoted 1 line

2. On unaligned DIO writes to reflinked CoW (not possible for atomic writes)

This path doesn't ever go through iomap - XFS catches that case before it calls into iomap, so it's not relevant to how iomap behaves w.r.t atomic IO.

Quoted 8 lines

So it anyways should never happen that XFS ever fallback to buffered-io for DIO atomic writes. Even today it does not fallback to buffered-io for non-atomic short DIO writes.

Quoted 4 lines

Quoted 3 lines

You pointed it right that it should be fixed in iomap. However do you think filesystems can still keep this as safety net (maybe no need of WARN_ON_ONCE).

Quoted 7 lines

Quoted 3 lines

I don't see any point in adding "impossible to hit" checks into filesystems just in case some core infrastructure has a bug introduced....

Yes, that is true for XFS. EXT4 however can return -ENOTBLK for short writes, though it should not happen for current atomic write case where we are only allowing for 1 fsblock.

Yes, but the -ENOTBLK error returned from ext4_iomap_end() if nothing was written does not get returned to ext4 from __iomap_dio_rw(). It is consumed by the iomap code:

	/* magic error code to fall back to buffered I/O */
        if (ret == -ENOTBLK) {
                wait_for_completion = true;
                ret = 0;
	}

This means that all the IO that was issued gets completed before returning to the caller and that's how the short write comes about.

-ENOTBLK is *not returned to the caller* on a short write -

iomap_dio_rw will return 0 (success). The caller then has to look at the iov_iter state to determine if the write was fully completed. This is exactly what the ext4 code currently does for all DIO writes, not just those that return -ENOTBLK.

Quoted 4 lines

I would still like to go with a WARN_ON_ONCE where we are calling ext4 buffered-io handling for DIO fallback writes. This is to catch any bugs even in future when we move to multi-fsblock case (until we have atomic write support for buffered-io).

Your choice, but please realise that it is not going to catch short atomic writes at all.

-Dave.

Signature · 2 lines

Dave Chinner
[email protected]

Filesystems like ext4 can submit writes in multiples of blocksizes. But we still can't allow the writes to be split. Hence let's check if the iomap_length() is same as iter->len or not.

This shouldn't affect XFS since it anyways checks for this in xfs_file_write_iter() to not support atomic write size request of more than FS blocksize.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
  fs/iomap/direct-io.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)

fs/iomap/direct-io.c +0 -0

index ed4764e3b8f0..1d33b4239b3e 100644

Quoted 21 lines · Ritesh Harjani (IBM)

Filesystems like ext4 can submit writes in multiples of blocksizes. But we still can't allow the writes to be split. Hence let's check if the iomap_length() is same as iter->len or not.

This shouldn't affect XFS since it anyways checks for this in xfs_file_write_iter() to not support atomic write size request of more than FS blocksize.

Signed-off-by Ritesh Harjani (IBM) <[email protected]>

---
  fs/iomap/direct-io.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)

fs/iomap/direct-io.c +0 -0

index ed4764e3b8f0..1d33b4239b3e 100644
@@ -306,7 +306,7 @@ static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
  	size_t copied = 0;
  	size_t orig_count;
 

fs/iomap/direct-io.c static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter, · line 310

Quoted 3 lines

fs/iomap/direct-io.c +1 -1

-	if (atomic && length != fs_block_size)
+	if (atomic && length != iter->len)
  		return -EINVAL;

Here you expect just one iter for an atomic write always.

In 6/6, you are saying that iomap does not allow an atomic write which covers unwritten and written extents, right?

But for writing a single fs block atomically, we don't mandate it to be in unwritten state. So there is a difference in behavior in writing a single FS block vs multiple FS blocks atomically.

So we have 3x choices, as I see:
a. add a check now in iomap that the extent is in written state (for an atomic write)
b. add extent zeroing code, as I was trying for originally c. document this peculiarity

Thanks,
John

Quoted 5 lines

fs/iomap/direct-io.c static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter, · line 310

Quoted 4 lines

fs/iomap/direct-io.c +1 -1

-	if (atomic && length != fs_block_size)
+	if (atomic && length != iter->len)
   		return -EINVAL;

Here you expect just one iter for an atomic write always.

Here we are lifting the limitation of iomap to support entire iter->len rather than just 1 fsblock.

Quoted 3 lines

In 6/6, you are saying that iomap does not allow an atomic write which covers unwritten and written extents, right?

No, it's not that. If FS does not provide a full mapping to iomap in ->iomap_begin then the writes will get split.

For atomic writes this
should not happen, hence the check in iomap above to return -EINVAL if mapped length does not match iter->len.

Quoted 3 lines

But for writing a single fs block atomically, we don't mandate it to be in unwritten state. So there is a difference in behavior in writing a single FS block vs multiple FS blocks atomically.

Same as mentioned above. We can't have atomic writes to get split. This patch is just lifting the restriction of iomap to allow more than blocksize but the mapped length should still meet iter->len, as otherwise the writes can get split.

Quoted 7 lines · Ritesh Harjani (IBM)

Quoted 5 lines

fs/iomap/direct-io.c static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter, · line 310

Quoted 4 lines

fs/iomap/direct-io.c +1 -1

-	if (atomic && length != fs_block_size)
+	if (atomic && length != iter->len)
   		return -EINVAL;

Here you expect just one iter for an atomic write always.

Here we are lifting the limitation of iomap to support entire iter->len rather than just 1 fsblock.

Sure

Quoted 5 lines

Quoted 3 lines

In 6/6, you are saying that iomap does not allow an atomic write which covers unwritten and written extents, right?

No, it's not that. If FS does not provide a full mapping to iomap in ->iomap_begin then the writes will get split.

but why would it provide multiple mapping?

Quoted 7 lines

For atomic writes this
should not happen, hence the check in iomap above to return -EINVAL if mapped length does not match iter->len.

Quoted 3 lines

But for writing a single fs block atomically, we don't mandate it to be in unwritten state. So there is a difference in behavior in writing a single FS block vs multiple FS blocks atomically.

Quoted 4 lines

Same as mentioned above. We can't have atomic writes to get split. This patch is just lifting the restriction of iomap to allow more than blocksize but the mapped length should still meet iter->len, as otherwise the writes can get split.

Sure, I get this. But I wonder why would we be getting multiple mappings? Why cannot the FS always provide a single mapping?

Quoted 6 lines

Quoted 4 lines

Same as mentioned above. We can't have atomic writes to get split. This patch is just lifting the restriction of iomap to allow more than blocksize but the mapped length should still meet iter->len, as otherwise the writes can get split.

Sure, I get this. But I wonder why would we be getting multiple mappings? Why cannot the FS always provide a single mapping?

FS can decide to split the mappings when it couldn't allocate a single
large mapping of the requested length. Could be due to -
- already allocated extent followed by EOF,
- already allocated extent followed by a hole
- already mapped extent followed by an extent of different type (e.g. written followed by unwritten or unwritten followed by written)

- delalloc (not delalloc since we invalidate respective page cache pages before doing DIO).
- fragmentation or ENOSPC - For ext4 bigalloc this will not happen since
we reserve the entire cluster. So we know there should be space. But I
am not sure how other filesystems might end up implementing this functionality.

Quoted 6 lines · Ritesh Harjani (IBM)

Quoted 6 lines

Quoted 4 lines

Same as mentioned above. We can't have atomic writes to get split. This patch is just lifting the restriction of iomap to allow more than blocksize but the mapped length should still meet iter->len, as otherwise the writes can get split.

Sure, I get this. But I wonder why would we be getting multiple mappings? Why cannot the FS always provide a single mapping?

Quoted 5 lines

FS can decide to split the mappings when it couldn't allocate a single
large mapping of the requested length. Could be due to -
- already allocated extent followed by EOF,
- already allocated extent followed by a hole
- already mapped extent followed by an extent of different type (e.g. written followed by unwritten or unwritten followed by written)

This is the sort of scenario which I am concerned with. This issue has been discussed at length for XFS forcealign support for atomic writes.

So far, the user can atomic write a single FS block regardless of whether the extent in which it would be part of is in written or unwritten state.

Now the rule will be to write multiple FS blocks atomically, all blocks need to be in same written or unwritten state.

This oddity at least needs to be documented.

Better yet would be to not have this restriction.

Quoted 4 lines

- delalloc (not delalloc since we invalidate respective page cache pages before doing DIO).
- fragmentation or ENOSPC - For ext4 bigalloc this will not happen since
we reserve the entire cluster. So we know there should be space. But I
am not sure how other filesystems might end up implementing this functionality.

Thanks,
John

Quoted 7 lines · John Garry

Quoted 6 lines · Ritesh Harjani (IBM)

Quoted 6 lines

Quoted 4 lines

Same as mentioned above. We can't have atomic writes to get split. This patch is just lifting the restriction of iomap to allow more than blocksize but the mapped length should still meet iter->len, as otherwise the writes can get split.

Sure, I get this. But I wonder why would we be getting multiple mappings? Why cannot the FS always provide a single mapping?

Quoted 8 lines

Quoted 5 lines

FS can decide to split the mappings when it couldn't allocate a single
large mapping of the requested length. Could be due to -
- already allocated extent followed by EOF,
- already allocated extent followed by a hole
- already mapped extent followed by an extent of different type (e.g. written followed by unwritten or unwritten followed by written)

This is the sort of scenario which I am concerned with. This issue has been discussed at length for XFS forcealign support for atomic writes.

extsize and forcealign is being worked for ext4 as well where we can add such support, sure.

Quoted 7 lines

So far, the user can atomic write a single FS block regardless of whether the extent in which it would be part of is in written or unwritten state.

Now the rule will be to write multiple FS blocks atomically, all blocks need to be in same written or unwritten state.

FS needs to ensure that the writes does not get torned. So for whatever reason FS splits the mapping then we need to return an -EINVAL error to not allow such writes to get torned. This patch just does that.

But I get your point. More below.

Quoted 2 lines

This oddity at least needs to be documented.

Got it. Yes, we can do that.

Quoted 1 line

Quoted 2 lines

Better yet would be to not have this restriction.

I haven't thought of a clever way where we don't have to zero out the rest of the unwritten mapping. With ext4 bigalloc since the entire cluster is anyway reserved - I was thinking if we can come up with a clever way for doing atomic writes to the entire user requested size w/o zeroing out.

Zeroing out the other unwritten extent is also a cost penalty to the user anyways.

So user will anyway will have to be made aware of not to attempt writes of fashion which can cause them such penalties.

As patch-6 mentions this is a base support for bs = ps systems for enabling atomic writes using bigalloc. For now we return -EINVAL when we can't allocate a continuous user requested mapping which means it won't support operations of types 8k followed by 16k.

Quoted 41 lines · Ritesh Harjani (IBM)

Quoted 7 lines · John Garry

Quoted 6 lines · Ritesh Harjani (IBM)

Quoted 6 lines

Quoted 4 lines

Same as mentioned above. We can't have atomic writes to get split. This patch is just lifting the restriction of iomap to allow more than blocksize but the mapped length should still meet iter->len, as otherwise the writes can get split.

Sure, I get this. But I wonder why would we be getting multiple mappings? Why cannot the FS always provide a single mapping?

Quoted 8 lines

Quoted 5 lines

FS can decide to split the mappings when it couldn't allocate a single
large mapping of the requested length. Could be due to -
- already allocated extent followed by EOF,
- already allocated extent followed by a hole
- already mapped extent followed by an extent of different type (e.g. written followed by unwritten or unwritten followed by written)

This is the sort of scenario which I am concerned with. This issue has been discussed at length for XFS forcealign support for atomic writes.

extsize and forcealign is being worked for ext4 as well where we can add such support, sure.

Quoted 7 lines

So far, the user can atomic write a single FS block regardless of whether the extent in which it would be part of is in written or unwritten state.

Now the rule will be to write multiple FS blocks atomically, all blocks need to be in same written or unwritten state.

FS needs to ensure that the writes does not get torned. So for whatever reason FS splits the mapping then we need to return an -EINVAL error to not allow such writes to get torned. This patch just does that.

But I get your point. More below.

Quoted 2 lines

This oddity at least needs to be documented.

Got it. Yes, we can do that.

Quoted 1 line

Quoted 8 lines

Quoted 2 lines

Better yet would be to not have this restriction.

I haven't thought of a clever way where we don't have to zero out the rest of the unwritten mapping. With ext4 bigalloc since the entire cluster is anyway reserved - I was thinking if we can come up with a clever way for doing atomic writes to the entire user requested size w/o zeroing out.

This following was main method which was being attempted:

https://lore.kernel.org/linux-fsdevel/[email protected]/

There were other ideas in different versions of the forcelign/xfs block atomic writes series.

Quoted 3 lines

Zeroing out the other unwritten extent is also a cost penalty to the user anyways.

Sure, unless we have a special inode flag to say "pre-zero the extent".

Quoted 8 lines

So user will anyway will have to be made aware of not to attempt writes of fashion which can cause them such penalties.

As patch-6 mentions this is a base support for bs = ps systems for enabling atomic writes using bigalloc. For now we return -EINVAL when we can't allocate a continuous user requested mapping which means it won't support operations of types 8k followed by 16k.

That's my least-preferred option.

I think better would be reject atomic writes that cover unwritten extents always - but that boat is about to sail...

Quoted 8 lines

Quoted 8 lines

So user will anyway will have to be made aware of not to attempt writes of fashion which can cause them such penalties.

As patch-6 mentions this is a base support for bs = ps systems for enabling atomic writes using bigalloc. For now we return -EINVAL when we can't allocate a continuous user requested mapping which means it won't support operations of types 8k followed by 16k.

Quoted 4 lines

That's my least-preferred option.

I think better would be reject atomic writes that cover unwritten extents always - but that boat is about to sail...

That's what this patch does.

For whatever reason if we couldn't allocate
a single contiguous region of requested size for atomic write, then we reject the request always, isn't it. Or maybe I didn't understand your comment.

If others prefer - we can maybe add such a check (e.g. ext4_dio_atomic_write_checks()) for atomic writes in ext4_dio_write_checks(), similar to how we detect overwrites case to decide whether we need a read v/s write semaphore.
So this can check if the user has a partially allocated extent for the user requested region and if yes, we can return -EINVAL from ext4_dio_write_iter() itself.

Quoted 1 line

I think this maybe better option than waiting until ->iomap_begin().

This might also bring all atomic write constraints to be checked in one place i.e. during ext4_file_write_iter() itself.

Quoted 8 lines · Ritesh Harjani (IBM)

Quoted 8 lines

Quoted 8 lines

So user will anyway will have to be made aware of not to attempt writes of fashion which can cause them such penalties.

As patch-6 mentions this is a base support for bs = ps systems for enabling atomic writes using bigalloc. For now we return -EINVAL when we can't allocate a continuous user requested mapping which means it won't support operations of types 8k followed by 16k.

Quoted 5 lines

Quoted 4 lines

That's my least-preferred option.

I think better would be reject atomic writes that cover unwritten extents always - but that boat is about to sail...

That's what this patch does.

Not really.

Currently we have 2x iomap restrictions:
a. mapping length must equal fs block size
b. bio created must equal total write size

This patch just says that the mapping length must equal total write size (instead of a.). So quite similar to b.

Quoted 3 lines

For whatever reason if we couldn't allocate
a single contiguous region of requested size for atomic write, then we reject the request always, isn't it. Or maybe I didn't understand your comment.

As the simplest example, for an atomic write to an empty file, there should only be a single mapping returned to iomap_dio_bio_iter() and that would be of IOMAP_UNWRITTEN type. And we don't reject that.

Quoted 7 lines

If others prefer - we can maybe add such a check (e.g. ext4_dio_atomic_write_checks()) for atomic writes in ext4_dio_write_checks(), similar to how we detect overwrites case to decide whether we need a read v/s write semaphore.
So this can check if the user has a partially allocated extent for the user requested region and if yes, we can return -EINVAL from ext4_dio_write_iter() itself.

Quoted 3 lines

Quoted 1 line

I think this maybe better option than waiting until ->iomap_begin().

This might also bring all atomic write constraints to be checked in one place i.e. during ext4_file_write_iter() itself.

Something like this can be done once we decide how atomic writing to regions which cover mixed unwritten and written extents is to be handled.

Quoted 24 lines · John Garry

Quoted 8 lines · Ritesh Harjani (IBM)

Quoted 8 lines

Quoted 8 lines

So user will anyway will have to be made aware of not to attempt writes of fashion which can cause them such penalties.

As patch-6 mentions this is a base support for bs = ps systems for enabling atomic writes using bigalloc. For now we return -EINVAL when we can't allocate a continuous user requested mapping which means it won't support operations of types 8k followed by 16k.

Quoted 5 lines

Quoted 4 lines

That's my least-preferred option.

I think better would be reject atomic writes that cover unwritten extents always - but that boat is about to sail...

That's what this patch does.

Not really.

Currently we have 2x iomap restrictions:
a. mapping length must equal fs block size
b. bio created must equal total write size

This patch just says that the mapping length must equal total write size (instead of a.). So quite similar to b.

Quoted 8 lines

Quoted 3 lines

For whatever reason if we couldn't allocate
a single contiguous region of requested size for atomic write, then we reject the request always, isn't it. Or maybe I didn't understand your comment.

As the simplest example, for an atomic write to an empty file, there should only be a single mapping returned to iomap_dio_bio_iter() and that would be of IOMAP_UNWRITTEN type. And we don't reject that.

Ok. Maybe this is what I am missing. Could you please help me understand why should such writes be rejected?

For e.g.
If FS could allocate a single contiguous IOMAP_UNWRITTEN extent of atomic write request size, that means -
1. FS will allocate an unwritten extent.
2. will do writes (using submit_bio) to the unwritten extent.
3. will do unwritten to written conversion.

It is ok if either of the above operations fail right? If (3) fails then the region will still be marked unwritten that means it will read zero (old contents). (2) can anyway fail and will not result into partial writes. (1) will anyway not result into any write whatsoever.

So we can never have a situation where there is partial writes leading to mix of old and new write contents right for such cases? Which is what the requirement of atomic/untorn write also is?

Sorry am I missing something here?

Quoted 7 lines

Quoted 7 lines

If others prefer - we can maybe add such a check (e.g. ext4_dio_atomic_write_checks()) for atomic writes in ext4_dio_write_checks(), similar to how we detect overwrites case to decide whether we need a read v/s write semaphore.
So this can check if the user has a partially allocated extent for the user requested region and if yes, we can return -EINVAL from ext4_dio_write_iter() itself.

Quoted 6 lines

Quoted 3 lines

Quoted 1 line

I think this maybe better option than waiting until ->iomap_begin().

This might also bring all atomic write constraints to be checked in one place i.e. during ext4_file_write_iter() itself.

Something like this can be done once we decide how atomic writing to regions which cover mixed unwritten and written extents is to be handled.

Mixed extent regions (written + unwritten) is a different case all together (which can lead to mix of old and new contents).

But here what I am suggesting is to add following constraint in case of ext4 with bigalloc -

"Writes to a region which already has partially allocated extent is not supported."

That means we will return -EINVAL if we detect above case in ext4_file_write_iter() and sure we can document this behavior.

In retrospect, I am not sure why we cannot add a constraint for atomic writes (e.g. for ext4 bigalloc) and reject such writes outright, instead of silently incurring a performance penalty by zeroing out the partial regions by allowing such write request.

-ritesh

Quoted 50 lines · Ritesh Harjani

Quoted 24 lines · John Garry

Quoted 8 lines · Ritesh Harjani (IBM)

Quoted 8 lines

Quoted 8 lines

So user will anyway will have to be made aware of not to attempt writes of fashion which can cause them such penalties.

As patch-6 mentions this is a base support for bs = ps systems for enabling atomic writes using bigalloc. For now we return -EINVAL when we can't allocate a continuous user requested mapping which means it won't support operations of types 8k followed by 16k.

Quoted 5 lines

Quoted 4 lines

That's my least-preferred option.

I think better would be reject atomic writes that cover unwritten extents always - but that boat is about to sail...

That's what this patch does.

Not really.

Currently we have 2x iomap restrictions:
a. mapping length must equal fs block size
b. bio created must equal total write size

This patch just says that the mapping length must equal total write size (instead of a.). So quite similar to b.

Quoted 8 lines

Quoted 3 lines

For whatever reason if we couldn't allocate
a single contiguous region of requested size for atomic write, then we reject the request always, isn't it. Or maybe I didn't understand your comment.

As the simplest example, for an atomic write to an empty file, there should only be a single mapping returned to iomap_dio_bio_iter() and that would be of IOMAP_UNWRITTEN type. And we don't reject that.

Ok. Maybe this is what I am missing. Could you please help me understand why should such writes be rejected?

For e.g.
If FS could allocate a single contiguous IOMAP_UNWRITTEN extent of atomic write request size, that means -
1. FS will allocate an unwritten extent.
2. will do writes (using submit_bio) to the unwritten extent.
3. will do unwritten to written conversion.

It is ok if either of the above operations fail right? If (3) fails then the region will still be marked unwritten that means it will read zero (old contents). (2) can anyway fail and will not result into partial writes. (1) will anyway not result into any write whatsoever.

Quoted 5 lines

So we can never have a situation where there is partial writes leading to mix of old and new write contents right for such cases? Which is what the requirement of atomic/untorn write also is?

Sorry am I missing something here?

I must be missing something; to perform an untorn write, two things must happen --

1. The kernel writes the data to the storage device, and the storage device either persists all of it, or throws back an error having persisted none of it.

2. If (1) completes successfully, all file mapping updates for the range written must be persisted, or an error is thrown back and none of them are persisted.

iomap doesn't have to know how the filesystem satisfies (2); it just has to create a single bio containing all data pages or it rejects the write.

Currently, it's an implementation detail that the XFS directio write ioend code processes the file mapping updates for the range written by walking every extent mapping for that range and issuing separate transactions for each mapping update. There's nothing that can restart the walk if it is interrupted. That's why XFS cannot support multi fsblock untorn writes to blocks with different status.

As I've said before, the most general solution to this would be to add a new log intent item that would track the "update all mappings in this file range" operation so that recovery could restart the walk. This is the most technically challenging, so we decided not to implement it until there is demand.

Having set aside the idea of redesigning ioend, the second-most general solution is pre-zeroing unwritten extents and holes so that ->iomap_begin implementations can present a single mapping to the bio constructor. Technically if there's only one unwritten extent or hole or cow, xfs can actually satisfy (2) because it only creates one transaction.

This gets me to the third and much less general solution -- only allow untorn writes if we know that the ioend only ever has to run a single transaction. That's why untorn writes are limited to a single fsblock for now -- it's a simple solution so that we can get our downstream customers to kick the tires and start on the next iteration instead of spending years on waterfalling.

Did you notice that in all of these cases, the capabilities of the filesystem's ioend processing determines the restrictions on the number and type of mappings that ->iomap_begin can give to iomap?

Now that we have a second system trying to hook up to the iomap support, it's clear to me that the restrictions on mappings are specific to each filesystem. Therefore, the iomap directio code should not impose restrictions on the mappings it receives unless they would prevent the creation of the single aligned bio.

Instead, xfs_direct_write_iomap_begin and ext4_iomap_begin should return EINVAL or something if they look at the file mappings and discover that they cannot perform the ioend without risking torn mapping updates. In the long run, ->iomap_begin is where this iomap->len <= iter->len check really belongs, but hold that thought.

For the multi fsblock case, the ->iomap_begin functions would have to check that only one metadata update would be necessary in the ioend. That's where things get murky, since ext4/xfs drop their mapping locks between calls to ->iomap_begin. So you'd have to check all the mappings for unsupported mixed state every time. Yuck.

It might be less gross to retain the restriction that iomap accepts only one mapping for the entire file range, like Ritesh has here.

Users
might be ok with us saying that you can't do a 16k atomic write to a region where you previously did an 8k write until you write the other 8k, even if someone has to write zeroes. Users might be ok with the kernel allowing multi-fsblock writes but only if the stars align.

But
to learn the answers to those questions, we have to put /something/ in the hands of our users.

For now (because we're already at -rc5), let's have xfs/ext4's ->write_iter implementations restrict atomic writes to a single fsblock, and get both merged into the kernel.

Let's defer the multi fsblock work
to 6.14, though I think we could take this patch.

Does that sound cool?

--D

Quoted 33 lines

Quoted 7 lines

Quoted 7 lines

If others prefer - we can maybe add such a check (e.g. ext4_dio_atomic_write_checks()) for atomic writes in ext4_dio_write_checks(), similar to how we detect overwrites case to decide whether we need a read v/s write semaphore.
So this can check if the user has a partially allocated extent for the user requested region and if yes, we can return -EINVAL from ext4_dio_write_iter() itself.

Quoted 6 lines

Quoted 3 lines

Quoted 1 line

I think this maybe better option than waiting until ->iomap_begin().

This might also bring all atomic write constraints to be checked in one place i.e. during ext4_file_write_iter() itself.

Something like this can be done once we decide how atomic writing to regions which cover mixed unwritten and written extents is to be handled.

Mixed extent regions (written + unwritten) is a different case all together (which can lead to mix of old and new contents).

But here what I am suggesting is to add following constraint in case of ext4 with bigalloc -

"Writes to a region which already has partially allocated extent is not supported."

That means we will return -EINVAL if we detect above case in ext4_file_write_iter() and sure we can document this behavior.

In retrospect, I am not sure why we cannot add a constraint for atomic writes (e.g. for ext4 bigalloc) and reject such writes outright, instead of silently incurring a performance penalty by zeroing out the partial regions by allowing such write request.

-ritesh

Quoted 115 lines · Darrick J. Wong

Quoted 50 lines · Ritesh Harjani

Quoted 24 lines · John Garry

Quoted 8 lines · Ritesh Harjani (IBM)

Quoted 8 lines

Quoted 8 lines

So user will anyway will have to be made aware of not to attempt writes of fashion which can cause them such penalties.

As patch-6 mentions this is a base support for bs = ps systems for enabling atomic writes using bigalloc. For now we return -EINVAL when we can't allocate a continuous user requested mapping which means it won't support operations of types 8k followed by 16k.

Quoted 5 lines

Quoted 4 lines

That's my least-preferred option.

I think better would be reject atomic writes that cover unwritten extents always - but that boat is about to sail...

That's what this patch does.

Not really.

Currently we have 2x iomap restrictions:
a. mapping length must equal fs block size
b. bio created must equal total write size

This patch just says that the mapping length must equal total write size (instead of a.). So quite similar to b.

Quoted 8 lines

Quoted 3 lines

For whatever reason if we couldn't allocate
a single contiguous region of requested size for atomic write, then we reject the request always, isn't it. Or maybe I didn't understand your comment.

As the simplest example, for an atomic write to an empty file, there should only be a single mapping returned to iomap_dio_bio_iter() and that would be of IOMAP_UNWRITTEN type. And we don't reject that.

Ok. Maybe this is what I am missing. Could you please help me understand why should such writes be rejected?

For e.g.
If FS could allocate a single contiguous IOMAP_UNWRITTEN extent of atomic write request size, that means -
1. FS will allocate an unwritten extent.
2. will do writes (using submit_bio) to the unwritten extent.
3. will do unwritten to written conversion.

It is ok if either of the above operations fail right? If (3) fails then the region will still be marked unwritten that means it will read zero (old contents). (2) can anyway fail and will not result into partial writes. (1) will anyway not result into any write whatsoever.

Quoted 5 lines

So we can never have a situation where there is partial writes leading to mix of old and new write contents right for such cases? Which is what the requirement of atomic/untorn write also is?

Sorry am I missing something here?

I must be missing something; to perform an untorn write, two things must happen --

1. The kernel writes the data to the storage device, and the storage device either persists all of it, or throws back an error having persisted none of it.

2. If (1) completes successfully, all file mapping updates for the range written must be persisted, or an error is thrown back and none of them are persisted.

iomap doesn't have to know how the filesystem satisfies (2); it just has to create a single bio containing all data pages or it rejects the write.

Currently, it's an implementation detail that the XFS directio write ioend code processes the file mapping updates for the range written by walking every extent mapping for that range and issuing separate transactions for each mapping update. There's nothing that can restart the walk if it is interrupted. That's why XFS cannot support multi fsblock untorn writes to blocks with different status.

As I've said before, the most general solution to this would be to add a new log intent item that would track the "update all mappings in this file range" operation so that recovery could restart the walk. This is the most technically challenging, so we decided not to implement it until there is demand.

Having set aside the idea of redesigning ioend, the second-most general solution is pre-zeroing unwritten extents and holes so that ->iomap_begin implementations can present a single mapping to the bio constructor. Technically if there's only one unwritten extent or hole or cow, xfs can actually satisfy (2) because it only creates one transaction.

This gets me to the third and much less general solution -- only allow untorn writes if we know that the ioend only ever has to run a single transaction. That's why untorn writes are limited to a single fsblock for now -- it's a simple solution so that we can get our downstream customers to kick the tires and start on the next iteration instead of spending years on waterfalling.

Did you notice that in all of these cases, the capabilities of the filesystem's ioend processing determines the restrictions on the number and type of mappings that ->iomap_begin can give to iomap?

Now that we have a second system trying to hook up to the iomap support, it's clear to me that the restrictions on mappings are specific to each filesystem. Therefore, the iomap directio code should not impose restrictions on the mappings it receives unless they would prevent the creation of the single aligned bio.

Instead, xfs_direct_write_iomap_begin and ext4_iomap_begin should return EINVAL or something if they look at the file mappings and discover that they cannot perform the ioend without risking torn mapping updates. In the long run, ->iomap_begin is where this iomap->len <= iter->len check really belongs, but hold that thought.

Quoted 6 lines

For the multi fsblock case, the ->iomap_begin functions would have to check that only one metadata update would be necessary in the ioend. That's where things get murky, since ext4/xfs drop their mapping locks between calls to ->iomap_begin. So you'd have to check all the mappings for unsupported mixed state every time. Yuck.

Thanks Darrick for taking time summarizing what all has been done and your thoughts here.

Quoted 2 lines

It might be less gross to retain the restriction that iomap accepts only one mapping for the entire file range, like Ritesh has here.

less gross :) sure.

I would like to think of this as, being less restrictive (compared to only allowing a single fsblock) by adding a constraint on the atomic write I/O request i.e.

"Atomic write I/O request to a region in a file is only allowed if that region has no partially allocated extents. Otherwise, the file system can fail the I/O operation by returning -EINVAL."

Essentially by adding this constraint to the I/O request, we are helping the user to prevent atomic writes from accidentally getting torned and also allowing multi-fsblock writes. So I still think that might be the right thing to do here or at least a better start. FS can later work on adding such support where we don't even need above such constraint on a given atomic write I/O request.

Quoted 9 lines

Users
might be ok with us saying that you can't do a 16k atomic write to a region where you previously did an 8k write until you write the other 8k, even if someone has to write zeroes. Users might be ok with the kernel allowing multi-fsblock writes but only if the stars align.

But
to learn the answers to those questions, we have to put /something/ in the hands of our users.

On this point, I think ext4 might already has those users who might be using atomic write characteristics of devices to do untorn writes. e.g.

In [1], Ted has talked about using bigalloc with ext4 for torn write prevention. [2] talks about using ext4 with bigalloc to prevent torn writes on aws cloud.

[1]: https://www.youtube.com/watch?v=gIeuiGg-_iw
[2]: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/configure-twp.html

My point being - Looks like the class of users who are using untorn writes to improve their database performances are already doing so even w/o any such interfaces being exposed to them (with ext4 bigalloc).

The current feature support of allowing atomic writes to only single fsblock might not be helpful to these users who can provide that feedback, who are using ext4 on bs = ps systems with bigalloc. But maybe let's wait and hear from them whether it is ok if -

"Atomic write I/O request to a region in a file is only allowed if that region has no partially allocated extents. Otherwise, the file system can fail the I/O operation by returning -EINVAL."

Quoted 4 lines

For now (because we're already at -rc5), let's have xfs/ext4's ->write_iter implementations restrict atomic writes to a single fsblock, and get both merged into the kernel.

Yes, I agree with the approach. I agree that we should get a consensus on this from folks.

Let me split this series up and address the review comments on patch [1-4]. Patch-5 & 6 can be worked once we have conclusion on this and can be eyed for 6.14.

Quoted 2 lines

Let's defer the multi fsblock work
to 6.14, though I think we could take this patch.

It's ok to consider this patch along with multi-fsblock work then i.e. for 6.14.

Quoted 4 lines

Does that sound cool?

--D

Thanks Darrick :)

-ritesh

Quoted 33 lines

Quoted 33 lines

Quoted 7 lines

Quoted 7 lines

If others prefer - we can maybe add such a check (e.g. ext4_dio_atomic_write_checks()) for atomic writes in ext4_dio_write_checks(), similar to how we detect overwrites case to decide whether we need a read v/s write semaphore.
So this can check if the user has a partially allocated extent for the user requested region and if yes, we can return -EINVAL from ext4_dio_write_iter() itself.

Quoted 6 lines

Quoted 3 lines

Quoted 1 line

I think this maybe better option than waiting until ->iomap_begin().

This might also bring all atomic write constraints to be checked in one place i.e. during ext4_file_write_iter() itself.

Something like this can be done once we decide how atomic writing to regions which cover mixed unwritten and written extents is to be handled.

Mixed extent regions (written + unwritten) is a different case all together (which can lead to mix of old and new contents).

But here what I am suggesting is to add following constraint in case of ext4 with bigalloc -

"Writes to a region which already has partially allocated extent is not supported."

That means we will return -EINVAL if we detect above case in ext4_file_write_iter() and sure we can document this behavior.

In retrospect, I am not sure why we cannot add a constraint for atomic writes (e.g. for ext4 bigalloc) and reject such writes outright, instead of silently incurring a performance penalty by zeroing out the partial regions by allowing such write request.

-ritesh

Quoted 46 lines · Ritesh Harjani

Quoted 23 lines

This gets me to the third and much less general solution -- only allow untorn writes if we know that the ioend only ever has to run a single transaction. That's why untorn writes are limited to a single fsblock for now -- it's a simple solution so that we can get our downstream customers to kick the tires and start on the next iteration instead of spending years on waterfalling.

Did you notice that in all of these cases, the capabilities of the filesystem's ioend processing determines the restrictions on the number and type of mappings that ->iomap_begin can give to iomap?

Now that we have a second system trying to hook up to the iomap support, it's clear to me that the restrictions on mappings are specific to each filesystem. Therefore, the iomap directio code should not impose restrictions on the mappings it receives unless they would prevent the creation of the single aligned bio.

Instead, xfs_direct_write_iomap_begin and ext4_iomap_begin should return EINVAL or something if they look at the file mappings and discover that they cannot perform the ioend without risking torn mapping updates. In the long run, ->iomap_begin is where this iomap->len <= iter->len check really belongs, but hold that thought.

Quoted 6 lines

For the multi fsblock case, the ->iomap_begin functions would have to check that only one metadata update would be necessary in the ioend. That's where things get murky, since ext4/xfs drop their mapping locks between calls to ->iomap_begin. So you'd have to check all the mappings for unsupported mixed state every time. Yuck.

Thanks Darrick for taking time summarizing what all has been done and your thoughts here.

Quoted 2 lines

It might be less gross to retain the restriction that iomap accepts only one mapping for the entire file range, like Ritesh has here.

less gross :) sure.

I would like to think of this as, being less restrictive (compared to only allowing a single fsblock) by adding a constraint on the atomic write I/O request i.e.

"Atomic write I/O request to a region in a file is only allowed if that region has no partially allocated extents. Otherwise, the file system can fail the I/O operation by returning -EINVAL."

Quoted 6 lines

Essentially by adding this constraint to the I/O request, we are helping the user to prevent atomic writes from accidentally getting torned and also allowing multi-fsblock writes. So I still think that might be the right thing to do here or at least a better start. FS can later work on adding such support where we don't even need above such constraint on a given atomic write I/O request.

On today's ext4 call, Ted and Ritesh and I realized that there's a bit more to it than this -- it's not possible to support untorn writes to a mix of written/(cow,unwritten) mappings even if they all point to the same physical space. If the system fails after the storage device commits the write but before any of the ioend processing is scheduled, a subsequent read of the previously written blocks will produce the new data, but reads to the other areas will produce the old contents (or zeroes, or whatever). That's a torn write.

Therefore, iomap ought to stick to requiring that ->iomap_begin returns a single iomap to cover the entire file range for the untorn write. For an unwritten extent, the post-recovery read will see either zeroes or the new contents; for a single-mapping COW it'll see old or new contents but not both.

Quoted 54 lines · Darrick J. Wong

Quoted 46 lines · Ritesh Harjani

Quoted 23 lines

This gets me to the third and much less general solution -- only allow untorn writes if we know that the ioend only ever has to run a single transaction. That's why untorn writes are limited to a single fsblock for now -- it's a simple solution so that we can get our downstream customers to kick the tires and start on the next iteration instead of spending years on waterfalling.

Did you notice that in all of these cases, the capabilities of the filesystem's ioend processing determines the restrictions on the number and type of mappings that ->iomap_begin can give to iomap?

Now that we have a second system trying to hook up to the iomap support, it's clear to me that the restrictions on mappings are specific to each filesystem. Therefore, the iomap directio code should not impose restrictions on the mappings it receives unless they would prevent the creation of the single aligned bio.

Instead, xfs_direct_write_iomap_begin and ext4_iomap_begin should return EINVAL or something if they look at the file mappings and discover that they cannot perform the ioend without risking torn mapping updates. In the long run, ->iomap_begin is where this iomap->len <= iter->len check really belongs, but hold that thought.

Quoted 6 lines

For the multi fsblock case, the ->iomap_begin functions would have to check that only one metadata update would be necessary in the ioend. That's where things get murky, since ext4/xfs drop their mapping locks between calls to ->iomap_begin. So you'd have to check all the mappings for unsupported mixed state every time. Yuck.

Thanks Darrick for taking time summarizing what all has been done and your thoughts here.

Quoted 2 lines

It might be less gross to retain the restriction that iomap accepts only one mapping for the entire file range, like Ritesh has here.

less gross :) sure.

I would like to think of this as, being less restrictive (compared to only allowing a single fsblock) by adding a constraint on the atomic write I/O request i.e.

"Atomic write I/O request to a region in a file is only allowed if that region has no partially allocated extents. Otherwise, the file system can fail the I/O operation by returning -EINVAL."

Quoted 6 lines

Essentially by adding this constraint to the I/O request, we are helping the user to prevent atomic writes from accidentally getting torned and also allowing multi-fsblock writes. So I still think that might be the right thing to do here or at least a better start. FS can later work on adding such support where we don't even need above such constraint on a given atomic write I/O request.

Quoted 8 lines

On today's ext4 call, Ted and Ritesh and I realized that there's a bit more to it than this -- it's not possible to support untorn writes to a mix of written/(cow,unwritten) mappings even if they all point to the same physical space. If the system fails after the storage device commits the write but before any of the ioend processing is scheduled, a subsequent read of the previously written blocks will produce the new data, but reads to the other areas will produce the old contents (or zeroes, or whatever). That's a torn write.

I'm *really* surprised that people are only realising that IO completion processing for atomic writes *must be atomic*.

This was the foundational constraint that the forced alignment proposal for XFS was intended to address. i.e. to prevent fs operations from violating atomic write IO constraints (e.g. punching sub-atomic write size holes in the file) so that the physical IO can be done without tearing and the IO completion processing that exposes the new data can be done atomically.

Quoted 5 lines

Therefore, iomap ought to stick to requiring that ->iomap_begin returns a single iomap to cover the entire file range for the untorn write. For an unwritten extent, the post-recovery read will see either zeroes or the new contents; for a single-mapping COW it'll see old or new contents but not both.

I'm pretty sure we enforced that in the XFS mapping implemention for atomic writes using forced alignment. i.e. we have to return a correctly aligned, contiguous mapping to iomap or we have to return -EINVAL to indicate atomic write mapping failed.

Yes, we can check this in iomap, but it's really the filesystem that has to implement and enforce it...

-Dave.

Signature · 2 lines

Dave Chinner
[email protected]