btrfs: add send/receive support for reading/writing compressed data

From: Omar Sandoval <[email protected]>

Commit e77fbf990316 ("btrfs: send: prepare for v2 protocol") added _BTRFS_SEND_C_MAX_V* macros equal to the maximum command number for the version plus 1, but as written this creates gaps in the number space. The maximum command number is currently 22, and __BTRFS_SEND_C_MAX_V1 is accordingly 23. But then __BTRFS_SEND_C_MAX_V2 is 24, suggesting that v2 has a command numbered 23, and __BTRFS_SEND_C_MAX is 25, suggesting that 23 and 24 are valid commands.

Instead, let's explicitly number all of the commands, attributes, and sentinel MAX constants.

Reviewed-by Sweet Tea Dorminy <[email protected]>

Signed-off-by Omar Sandoval <[email protected]>

---
 fs/btrfs/send.c |   4 +-
 fs/btrfs/send.h | 106 ++++++++++++++++++++++++------------------------
 2 files changed, 54 insertions(+), 56 deletions(-)

fs/btrfs/send.c +2 -2

index 6d36dee1505f..9363f625fa17 100644
@@ -326,8 +326,8 @@ __maybe_unused
 static bool proto_cmd_ok(const struct send_ctx *sctx, int cmd)
 {
 	switch (sctx->proto) {
-	case 1:	 return cmd < __BTRFS_SEND_C_MAX_V1;
-	case 2:	 return cmd < __BTRFS_SEND_C_MAX_V2;
+	case 1:	 return cmd <= BTRFS_SEND_C_MAX_V1;
+	case 2:	 return cmd <= BTRFS_SEND_C_MAX_V2;
 	default: return false;
 	}
 }

fs/btrfs/send.h +3 -5

index 08602fdd600a..67721e0281ba 100644
@@ -46,84 +46,82 @@ struct btrfs_tlv_header {
 
 /* commands */
 enum btrfs_send_cmd {
-	BTRFS_SEND_C_UNSPEC,
+	BTRFS_SEND_C_UNSPEC = 0,
 
 	/* Version 1 */
-	BTRFS_SEND_C_SUBVOL,
-	BTRFS_SEND_C_SNAPSHOT,
+	BTRFS_SEND_C_SUBVOL = 1,
+	BTRFS_SEND_C_SNAPSHOT = 2,
 
-	BTRFS_SEND_C_MKFILE,
-	BTRFS_SEND_C_MKDIR,

From: Omar Sandoval <[email protected]>

This adds the definitions of the new commands for send stream version 2 and their respective attributes: fallocate, FS_IOC_SETFLAGS (a.k.a. chattr), and encoded writes. It also documents two changes to the send stream format in v2: the receiver shouldn't assume a maximum command size, and the DATA attribute is encoded differently to allow for writes larger than 64k. These will be implemented in subsequent changes, and then the ioctl will accept the new version and flag.

Reviewed-by Josef Bacik <[email protected]>

Reviewed-by Sweet Tea Dorminy <[email protected]>

Signed-off-by Omar Sandoval <[email protected]>

---
 fs/btrfs/send.c            |  2 +-
 fs/btrfs/send.h            | 40 ++++++++++++++++++++++++++++++++++----
 include/uapi/linux/btrfs.h |  7 +++++++
 3 files changed, 44 insertions(+), 5 deletions(-)

fs/btrfs/send.c +1 -1

index 9363f625fa17..1f141de3a7d6 100644
@@ -7459,7 +7459,7 @@ long btrfs_ioctl_send(struct inode *inode, struct btrfs_ioctl_send_args *arg)
 
 	sctx->clone_roots_cnt = arg->clone_sources_count;
 
-	sctx->send_max_size = BTRFS_SEND_BUF_SIZE;
+	sctx->send_max_size = BTRFS_SEND_BUF_SIZE_V1;
 	sctx->send_buf = kvmalloc(sctx->send_max_size, GFP_KERNEL);
 	if (!sctx->send_buf) {
 		ret = -ENOMEM;

fs/btrfs/send.h +5 -1

index 67721e0281ba..805d8095209a 100644
@@ -12,7 +12,11 @@
 #define BTRFS_SEND_STREAM_MAGIC "btrfs-stream"
 #define BTRFS_SEND_STREAM_VERSION 1
 
-#define BTRFS_SEND_BUF_SIZE SZ_64K
+/*
+ * In send stream v1, no command is larger than 64k. In send stream v2, no limit
+ * should be assumed.
+ */
+#define BTRFS_SEND_BUF_SIZE_V1 SZ_64K
 
 enum btrfs_tlv_type {
 	BTRFS_TLV_U8,

fs/btrfs/send.h +3 -1

+	BTRFS_SEND_C_ENCODED_WRITE = 25,
+	BTRFS_SEND_C_MAX_V2 = 25,
 
 	/* End */
-	BTRFS_SEND_C_MAX = 22,
+	BTRFS_SEND_C_MAX = 25,
 };

Quoted 52 lines · Omar Sandoval

From: Omar Sandoval <[email protected]>

This adds the definitions of the new commands for send stream version 2 and their respective attributes: fallocate, FS_IOC_SETFLAGS (a.k.a. chattr), and encoded writes. It also documents two changes to the send stream format in v2: the receiver shouldn't assume a maximum command size, and the DATA attribute is encoded differently to allow for writes larger than 64k. These will be implemented in subsequent changes, and then the ioctl will accept the new version and flag.

Reviewed-by Josef Bacik <[email protected]>

Reviewed-by Sweet Tea Dorminy <[email protected]>

Signed-off-by Omar Sandoval <[email protected]>

---
 fs/btrfs/send.c            |  2 +-
 fs/btrfs/send.h            | 40 ++++++++++++++++++++++++++++++++++----
 include/uapi/linux/btrfs.h |  7 +++++++
 3 files changed, 44 insertions(+), 5 deletions(-)

fs/btrfs/send.c +1 -1

index 9363f625fa17..1f141de3a7d6 100644
@@ -7459,7 +7459,7 @@ long btrfs_ioctl_send(struct inode *inode, struct btrfs_ioctl_send_args *arg)
 
 	sctx->clone_roots_cnt = arg->clone_sources_count;
 
-	sctx->send_max_size = BTRFS_SEND_BUF_SIZE;
+	sctx->send_max_size = BTRFS_SEND_BUF_SIZE_V1;
 	sctx->send_buf = kvmalloc(sctx->send_max_size, GFP_KERNEL);
 	if (!sctx->send_buf) {
 		ret = -ENOMEM;

fs/btrfs/send.h +5 -1

index 67721e0281ba..805d8095209a 100644
@@ -12,7 +12,11 @@
 #define BTRFS_SEND_STREAM_MAGIC "btrfs-stream"
 #define BTRFS_SEND_STREAM_VERSION 1
 
-#define BTRFS_SEND_BUF_SIZE SZ_64K
+/*
+ * In send stream v1, no command is larger than 64k. In send stream v2, no limit
+ * should be assumed.
+ */
+#define BTRFS_SEND_BUF_SIZE_V1 SZ_64K
 
 enum btrfs_tlv_type {
 	BTRFS_TLV_U8,

@@ -80,16 +84,20 @@ enum btrfs_send_cmd {
 	BTRFS_SEND_C_MAX_V1 = 22,
 

fs/btrfs/send.h enum btrfs_send_cmd { · line 88

Quoted 4 lines

fs/btrfs/send.h +2 -1

 	/* Version 2 */
-	BTRFS_SEND_C_MAX_V2 = 22,
+	BTRFS_SEND_C_FALLOCATE = 23,
+	BTRFS_SEND_C_SETFLAGS = 24,

Do you have patches that implement the fallocate modes and setflags? I don't see it in this patchset.

The setflags should be switched to
something closer to the recent refactoring that unifies all the flags/attrs to fileattr. I have a prototype patch for that, comparing the inode flags in the same way as file mode, the tricky part is on the receive side how to apply them correctly. On the sending side it's simple though.

Quoted 11 lines · David Sterba

fs/btrfs/send.h enum btrfs_send_cmd { · line 88

Quoted 7 lines · Omar Sandoval

fs/btrfs/send.h +2 -1

@@ -80,16 +84,20 @@ enum btrfs_send_cmd {
 	BTRFS_SEND_C_MAX_V1 = 22,
 
 	/* Version 2 */
-	BTRFS_SEND_C_MAX_V2 = 22,
+	BTRFS_SEND_C_FALLOCATE = 23,
+	BTRFS_SEND_C_SETFLAGS = 24,

Do you have patches that implement the fallocate modes and setflags? I don't see it in this patchset.

Nope, as discussed before, in order to keep the patch series managable, this series adds the definitions and receive support for fallocate and setflags, but leaves the send side to be implemented at a later time.

I implemented fallocate for send back in 2019:
https://github.com/osandov/linux/commits/btrfs-send-v2. It passed some basic testing back then, but it'd need a big rebase and more testing.

Quoted 6 lines

The setflags should be switched to
something closer to the recent refactoring that unifies all the flags/attrs to fileattr. I have a prototype patch for that, comparing the inode flags in the same way as file mode, the tricky part is on the receive side how to apply them correctly. On the sending side it's simple though.

The way this series documents (and implements in receive) BTRFS_SEND_C_SETFLAGS is that it's a simple call to FS_IOC_SETFLAGS with given flags. I don't think this is affected by the change to fileattr, unless I'm misunderstanding.

This is in line with the other commands being straightforward system calls, but it does mean that the sending side has to deal with the complexities of an immutable or append-only file being modified between incremental sends (by temporarily clearing the flag), and of inherited flags (e.g., a COW file inside of a NOCOW directory).

I suppose it'd
also be possible to have SETFLAGS define the final flags and leave it up to receive to make that happen by temporarily setting/clearing flags as necessary, but that is a bit inconsistent with how we've handled other commands.

Quoted 13 lines · Omar Sandoval

Quoted 11 lines · David Sterba

fs/btrfs/send.h enum btrfs_send_cmd { · line 88

Quoted 7 lines · Omar Sandoval

fs/btrfs/send.h +2 -1

@@ -80,16 +84,20 @@ enum btrfs_send_cmd {
 	BTRFS_SEND_C_MAX_V1 = 22,
 
 	/* Version 2 */
-	BTRFS_SEND_C_MAX_V2 = 22,
+	BTRFS_SEND_C_FALLOCATE = 23,
+	BTRFS_SEND_C_SETFLAGS = 24,

Do you have patches that implement the fallocate modes and setflags? I don't see it in this patchset.

Quoted 7 lines

Nope, as discussed before, in order to keep the patch series managable, this series adds the definitions and receive support for fallocate and setflags, but leaves the send side to be implemented at a later time.

I implemented fallocate for send back in 2019:
https://github.com/osandov/linux/commits/btrfs-send-v2. It passed some basic testing back then, but it'd need a big rebase and more testing.

The patches in the branch are partially cleanups and preparatory work, so at least avoiding sending the holes would be nice to have for v2 as it was one of the first bugs reported. The falllocate modes seem to be easy. The rest is about the versioning infrastructure that we already have merged.

Quoted 11 lines

Quoted 6 lines

The setflags should be switched to
something closer to the recent refactoring that unifies all the flags/attrs to fileattr. I have a prototype patch for that, comparing the inode flags in the same way as file mode, the tricky part is on the receive side how to apply them correctly. On the sending side it's simple though.

The way this series documents (and implements in receive) BTRFS_SEND_C_SETFLAGS is that it's a simple call to FS_IOC_SETFLAGS with given flags. I don't think this is affected by the change to fileattr, unless I'm misunderstanding.

The SETFLAGS ioctls are obsolete and I don't want to make them part of the protocol defition because the bit namespace contains flags we don't have implemented or are not releated to anything in btrfs.

https://elixir.bootlin.com/linux/latest/source/include/uapi/linux/fs.h#L220

It's basically just naming and specifying what exactly is the value so we should pick the most recent interface name that superseded SETFLAGS and the XFLAGS.

Quoted 5 lines

This is in line with the other commands being straightforward system calls, but it does mean that the sending side has to deal with the complexities of an immutable or append-only file being modified between incremental sends (by temporarily clearing the flag), and of inherited flags (e.g., a COW file inside of a NOCOW directory).

Yeah the receiving side needs to understand the constraints of the flags, it has only the information about the final state and not the order in which the flags get applied.

Quoted 9 lines · David Sterba

The SETFLAGS ioctls are obsolete and I don't want to make them part of the protocol defition because the bit namespace contains flags we don't have implemented or are not releated to anything in btrfs.

https://elixir.bootlin.com/linux/latest/source/include/uapi/linux/fs.h#L220

It's basically just naming and specifying what exactly is the value so we should pick the most recent interface name that superseded SETFLAGS and the XFLAGS.

This is the situation with FS_IOC_SETFLAGS, FS_IOC_FSSETXATTR, and fileattr as I understand it. Please correct me if I'm wrong:

- FS_IOC_SETFLAGS originally came from ext4 and was added to Btrfs very early on (commit 6cbff00f4632 ("Btrfs: implement FS_IOC_GETFLAGS/SETFLAGS/GETVERSION")).
- FS_IOC_FSSETXATTR originally came from XFS and was added to Btrfs a few years ago (in commit 025f2121488e ("btrfs: add FS_IOC_FSSETXATTR ioctl")).
- The two ioctls allow setting some of the same flags (e.g., IMMUTABLE, APPEND), but some are only supported by SETFLAGS (e.g., NOCOW) and some are only supported by FSSETXATTR (none of these are supported by Btrfs, however).
- fileattr is a recent VFS interface that is used to implement those two ioctls. It basically passes through the arguments for whichever ioctl was called and translates the equivalent flags between the two ioctls. It is not a new UAPI and doesn't have its own set of flags.

Is there another new UAPI that I'm missing that obsoletes SETFLAGS?

I see your point about the irrelevant flags in SETFLAGS, however. Is your suggestion to have our own send protocol-specific set of flags that we translate to whatever ioctl we need to make?

Quoted 9 lines

Quoted 5 lines

This is in line with the other commands being straightforward system calls, but it does mean that the sending side has to deal with the complexities of an immutable or append-only file being modified between incremental sends (by temporarily clearing the flag), and of inherited flags (e.g., a COW file inside of a NOCOW directory).

Yeah the receiving side needs to understand the constraints of the flags, it has only the information about the final state and not the order in which the flags get applied.

If the sender only tells the receiver what the final flags are, then yes, the receiver would need to deal with, e.g., temporarily clearing and resetting flags. The way I envisioned it was that the sender would instead send commands for those intermediate flag operations. E.g., if the incremental send requires writing some data to a file that is immutable in both the source and the parent subvolume, the sender could send commands to: clear the immutable flag, write the data, set the immutable flag. This is a lot like the orphan renaming that you mentioned.

If we want to have receive handle the intermediate states instead, then I would like to postpone SETFLAGS (or whatever we call it) to send protocol v3, since it'll be very tricky to get right and we can't add it to the protocol without having an implementation in the receiver.

On the other hand, if send takes care of the intermediate states and receive just has to blindly apply the flags, then we can add SETFLAGS to the protocol and receive now and implement it in send later. That is exactly what this patch series does.

I'm fine with either of those paths forward, but I don't want to block the compressed send/receive on SETFLAGS or fallocate.

Quoted 29 lines · Omar Sandoval

Quoted 9 lines · David Sterba

The SETFLAGS ioctls are obsolete and I don't want to make them part of the protocol defition because the bit namespace contains flags we don't have implemented or are not releated to anything in btrfs.

https://elixir.bootlin.com/linux/latest/source/include/uapi/linux/fs.h#L220

It's basically just naming and specifying what exactly is the value so we should pick the most recent interface name that superseded SETFLAGS and the XFLAGS.

This is the situation with FS_IOC_SETFLAGS, FS_IOC_FSSETXATTR, and fileattr as I understand it. Please correct me if I'm wrong:

- FS_IOC_SETFLAGS originally came from ext4 and was added to Btrfs very early on (commit 6cbff00f4632 ("Btrfs: implement FS_IOC_GETFLAGS/SETFLAGS/GETVERSION")).
- FS_IOC_FSSETXATTR originally came from XFS and was added to Btrfs a few years ago (in commit 025f2121488e ("btrfs: add FS_IOC_FSSETXATTR ioctl")).
- The two ioctls allow setting some of the same flags (e.g., IMMUTABLE, APPEND), but some are only supported by SETFLAGS (e.g., NOCOW) and some are only supported by FSSETXATTR (none of these are supported by Btrfs, however).
- fileattr is a recent VFS interface that is used to implement those two ioctls. It basically passes through the arguments for whichever ioctl was called and translates the equivalent flags between the two ioctls. It is not a new UAPI and doesn't have its own set of flags.

Quoted 1 line

Is there another new UAPI that I'm missing that obsoletes SETFLAGS?

That was supposed to be FSSETXATTR, new flags have appeared there, the reason for btrfs was to allow the FS_XFLAG_DAX bit as people are were working on the DAX support, and potentially other bits like FS_XFLAG_NOSYMLINKS or FS_XFLAG_NODEFRAG. Or new flags that we want to be able to set, NODATASUM for example.

Quoted 3 lines

I see your point about the irrelevant flags in SETFLAGS, however. Is your suggestion to have our own send protocol-specific set of flags that we translate to whatever ioctl we need to make?

Yes, that's the idea, the flags are not protocol-specific but rather btrfs-specific, ie we want to support namely the bits that btrfs inodes can have.

Quoted 10 lines

Quoted 9 lines

Quoted 5 lines

This is in line with the other commands being straightforward system calls, but it does mean that the sending side has to deal with the complexities of an immutable or append-only file being modified between incremental sends (by temporarily clearing the flag), and of inherited flags (e.g., a COW file inside of a NOCOW directory).

Yeah the receiving side needs to understand the constraints of the flags, it has only the information about the final state and not the order in which the flags get applied.

Quoted 9 lines

If the sender only tells the receiver what the final flags are, then yes, the receiver would need to deal with, e.g., temporarily clearing and resetting flags. The way I envisioned it was that the sender would instead send commands for those intermediate flag operations. E.g., if the incremental send requires writing some data to a file that is immutable in both the source and the parent subvolume, the sender could send commands to: clear the immutable flag, write the data, set the immutable flag. This is a lot like the orphan renaming that you mentioned.

I see, so the question is where do we want to put the logic. I'd go with userspace as lots of things are easier there, eg. maitaining some intermediate state or delayed application of bits/flags.

Quoted 4 lines

If we want to have receive handle the intermediate states instead, then I would like to postpone SETFLAGS (or whatever we call it) to send protocol v3, since it'll be very tricky to get right and we can't add it to the protocol without having an implementation in the receiver.

Yeah it would be tricky to generate the sequence right, while if it's on the receiving side we can simply ignore/report it or implement a subset where we know how to apply (eg. immutable) and don't need to postpone it.

Quoted 4 lines

On the other hand, if send takes care of the intermediate states and receive just has to blindly apply the flags, then we can add SETFLAGS to the protocol and receive now and implement it in send later. That is exactly what this patch series does.

It adds a command to the protocol but does not outline the plan how to use it, not counting this discussion.

Quoted 2 lines

I'm fine with either of those paths forward, but I don't want to block the compressed send/receive on SETFLAGS or fallocate.

I get that you care only about the encoded write, but I don't want to rev protocol every few releases because we did not bother to implement something we know is missing in the protocol. Anyway, encoded write will be in v2 scheduled for 5.20 and I'll implement the rest plus will have a look at your fallocate patches.