We're trying to get a U32 for the attributes, but the kernel sends a U64
(which is correct as we store attributes in a u64 flags field of the
inode). This makes anyone trying to receive a v2 send stream to fail with:
ERROR: invalid size for attribute, expected = 4, got = 8
We actually recently got such a report of someone using send stream v2 and
getting such failure. See the Link tag below.
Link: https://lore.kernel.org/linux-btrfs/6cb11fa5-c60d-e65b-0295-301a694e66ad@inbox.ru/
Fixes: 7a6fb356dc ("btrfs-progs: receive: process setflags ioctl commands")
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Process an enable_verity cmd by running the enable verity ioctl on the
file. Since enabling verity denies write access to the file, it is
important that we don't have any open write file descriptors.
This also revs the send stream format to version 3 with no format
changes besides the new commands and attributes. This version is not
finalized and commands may change, also this needs to be synchronized
with any kernel changes.
Note: the build is conditional on the header linux/fsverity.h
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
The tool IWYU (include what you use) suggests to remove and add some
includes. This is only partial to avoid accidental build breakage, the
includes are entangled and will have to be cleaned in the future again.
Signed-off-by: David Sterba <dsterba@suse.com>
The preferred order:
- system headers
- standard headers
- libraries
- kernel library
- kernel shared
- common headers
- other tools
- own headers
Signed-off-by: David Sterba <dsterba@suse.com>
The initial proposal for file attributes was built on simply doing
SETFLAGS but this builds on an old and non-extensible interface that has
no direct mapping for all inode flags. There's a unified interface
fileattr that covers file attributes and xflags, it should be possible
to add new bits.
On the protocol level the value is copied as-is in the original inode
but this does not provide enough information how to apply the bits on
the receiving side. Eg. IMMUTABLE flag prevents any changes to the file
and has to be handled manually.
The receiving side does not apply the bits yet, only parses it from the
stream.
Signed-off-by: David Sterba <dsterba@suse.com>
In send stream v2, send can emit a command for setting inode flags via
the setflags ioctl. Pass the flags attribute through to the ioctl call
in receive.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
Send stream v2 can emit fallocate commands, so receive must support them
as well. The implementation simply passes along the arguments to the
syscall. Note that mode is encoded as a u32 in send stream but fallocate
takes an int, so there is a unsigned->signed conversion there.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
Add a new btrfs_send_op and support for both dumping and proper receive
processing which does actual encoded writes.
Encoded writes are only allowed on a file descriptor opened with an
extra flag that allows encoded writes, so we also add support for this
flag when opening or reusing a file for writing.
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
The new format privileges the BTRFS_SEND_A_DATA attribute by
guaranteeing it will always be the last attribute in any command that
needs it, and by implicitly encoding the data length as the difference
between the total command length in the command header and the sizes of
the rest of the attributes (and of course the tlv_type identifying the
DATA attribute). To parse the new stream, we must read the tlv_type and
if it is not DATA, we proceed normally, but if it is DATA, we don't
parse a tlv_len but simply compute the length.
In addition, we add some bounds checking when parsing each chunk of
data, as well as for the tlv_len itself.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
In send stream v2, write commands can now be an arbitrary size. For that
reason, we can no longer allocate a fixed array in sctx for read_cmd.
Instead, read_cmd dynamically allocates sctx->read_buf. To avoid
needless reallocations, we reuse read_buf between read_cmd calls by also
keeping track of the size of the allocated buffer in sctx->read_buf_sz.
We do the first allocation of the old default size at the start of
processing the stream, and we only reallocate if we encounter a command
that needs a larger buffer.
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
An encoded extent can be up to 128K in length, which exceeds the largest
value expressible by the current send stream format's 16 bit tlv_len
field. Since encoded writes cannot be split into multiple writes by
btrfs send, the send stream format must change to accommodate encoded
writes.
Supporting this changed format requires retooling how we store the
commands we have processed. We currently store pointers to the struct
btrfs_tlv_headers in the command buffer. This is not sufficient to
represent the new BTRFS_SEND_A_DATA format. Instead, parse the attribute
headers and store them in a new struct btrfs_send_attribute which has a
32bit length field. This is transparent to users of the various TLV_GET
macros.
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
Kernel emits inode number for all mkfile/mkdir/... commands but the
receive part does not pass it to the callbacks. At least document that
and read it from the stream in case we'd like to use it in the future.
Signed-off-by: David Sterba <dsterba@suse.com>
The header contains the protocol definitions and is almost exactly the
same as the kernel version, move it to the proper directory.
Signed-off-by: David Sterba <dsterba@suse.com>
To use optimized CRC implementation, the input buffer must be
unsigned long aligned. btrfs receive calculates checksum based on
read_buf, including btrfs_cmd_header (with zeroed CRC field)
and command content.
Reorder the buffer to the beginning of the structure and force the
alignment to 64, this should be cacheline friendly and could speed up
the data transfers.
Interesting parts from the report:
Sending host:
Fedora 33
AMD ThreadRipper 1920X - 128GB RAM
2x10GBit Ethernet, bonded
MegaRaid 9270
6x16TB Seagate Exos in RAID5
Receiving host:
Fedora 33
Intel i3-7300 - HT enabled - 32GB RAM
10GBit Ethernet, single connection
MegaRaid 9260
12x8TB WD NAS drives in RAID5
The 2 hosts are connected to the same 10G switch. The sender could definitely
saturate a 10GBit link. The practically achievable writes on the backup host
would be lower, but still at least 400MB/s. The file system contains mostly
large files of 1GB+, so there is little meta-data.
With btrfs send/receive I'm getting a steady transfer rate of 60MB/s. The copy
has been running for a little over 5 days now, having only transferred some
25TB. This is way too slow for this setup.
Analyzing resource usage, the sender side is fine, both the btrfs send and the
corresponding ssh process only use about 10-10% CPU, which on a 24 threaded
machine is virtually nothing. However, the receiver is running with a load of
~2.6, with the sshd using 30-50% CPU and the btrfs receive a further 60-70%.
The rest of the load comes from IO wait. So the bottleneck is the btrfs receive
clearly.
Issue: #324
Signed-off-by: Sheng Mao <shngmao@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.com>
