| Message ID | cover.1597994106.git.osandov@osandov.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | btrfs: implement send/receive of compressed extents without decompressing | expand |
On Fri, Aug 21, 2020 at 12:39:50AM -0700, Omar Sandoval wrote: > Protocol Updates > ================ > > This series makes some changes to the send stream protocol beyond adding > the encoded write command/attributes and bumping the version. Namely, v1 > has a 64k limit on the size of a write due to the 16-bit attribute > length. This is not enough for encoded writes, as compressed extents may > be up to 128k and cannot be split up. To address this, the > BTRFS_SEND_A_DATA is treated specially in v2: its length is implicitly > the remaining length of the command (which has a 32-bit length). This > was the last bad of the options I considered. > > There are other commands that we've been wanting to add to the protocol: > fallocate and FS_IOC_SETFLAGS. This series reserves their command and > attribute numbers but does not implement kernel support for emitting > them. However, it does implement support in receive for them, so the > kernel can start emitting those whenever we get around to implementing > them. Can you please outline the protocol changes (as a bullet list) and eventually cross-ref with items https://btrfs.wiki.kernel.org/index.php/Design_notes_on_Send/Receive#Send_stream_v2_draft I'd like to know which and why you did not implement. The decision here is between get v2 out with most desired options and rev v3 later with the rest, or do v2 as complete as possible.
On Mon, Aug 24, 2020 at 09:57:55PM +0200, David Sterba wrote: > On Fri, Aug 21, 2020 at 12:39:50AM -0700, Omar Sandoval wrote: > > Protocol Updates > > ================ > > > > This series makes some changes to the send stream protocol beyond adding > > the encoded write command/attributes and bumping the version. Namely, v1 > > has a 64k limit on the size of a write due to the 16-bit attribute > > length. This is not enough for encoded writes, as compressed extents may > > be up to 128k and cannot be split up. To address this, the > > BTRFS_SEND_A_DATA is treated specially in v2: its length is implicitly > > the remaining length of the command (which has a 32-bit length). This > > was the last bad of the options I considered. > > > > There are other commands that we've been wanting to add to the protocol: > > fallocate and FS_IOC_SETFLAGS. This series reserves their command and > > attribute numbers but does not implement kernel support for emitting > > them. However, it does implement support in receive for them, so the > > kernel can start emitting those whenever we get around to implementing > > them. > > Can you please outline the protocol changes (as a bullet list) and > eventually cross-ref with items > https://btrfs.wiki.kernel.org/index.php/Design_notes_on_Send/Receive#Send_stream_v2_draft > > I'd like to know which and why you did not implement. The decision here > is between get v2 out with most desired options and rev v3 later with > the rest, or do v2 as complete as possible. The short version is that I didn't implement the kernel side of any of those :) the RWF_ENCODED series + this series is already big, and I didn't want to make it even bigger. I figured updating the protocol/receive now and doing the kernel side later was a good compromise (rather than doing a huge code dump or constantly bumping the protocol version). Is there some reason you don't like this approach? I'm of course happy to go about this in whatever way you think is best. Here's a breakdown of the list from the wiki: * Send extent holes, send preallocated extents: both require fallocate. Boris implemented the receive side. I have some old patches implementing the send side [1], but they're a largish rework of extent tracking in send. * Extent clones within one file: as far as I can tell, this is already possible with v1, it just sends redundant file paths. * Send otime for inodes: the consensus when I posted patches to enable this [2] was that we don't want this after all. * Send file flags (FS_IOC_GETFLAGS/FS_IOC_SETFLAGS): again, Boris implemented the receive side. I previously took a stab at the send side, but it's really annoying because of all of the interactions between directory inheritance, writes vs. NOCOW/append-only/immutable, etc. It's do-able, it would just take a lot of care. * Optionally send owner/group as strings: this one I wasn't aware of. * "block device is not sent over the stream": I don't know what this is referring to. It looks like we send block device nodes with mknod. In my opinion, fallocate support is the most important, SETFLAGS would be good but is a lot of effort, and the rest are nice-to-have. Let me know how you'd like me to go about this. 1: https://github.com/osandov/linux/commits/btrfs-send-v2 2: https://lore.kernel.org/linux-btrfs/cover.1550136164.git.osandov@fb.com/
On Fri, Aug 21, 2020 at 12:39:50AM -0700, Omar Sandoval wrote: > Omar Sandoval (9): > btrfs: send: get rid of i_size logic in send_write() > btrfs: send: avoid copying file data > btrfs: send: use btrfs_file_extent_end() in send_write_or_clone() > btrfs: add send_stream_version attribute to sysfs For the record, I'll add 1-4 to misc-next.
This series uses the interface added in "fs: interface for directly reading/writing compressed data" to send and receive compressed data without wastefully decompressing and recompressing it. It does so by 1. Bumping the send stream protocol version to 2. 2. Adding a new command, BTRFS_SEND_C_ENCODED_WRITE, and its associated attributes that indicates a write using the new encoded I/O interface. 3. Sending compressed extents with BTRFS_SEND_C_ENCODED_WRITE when requested by the user. 4. Falling back to decompressing and writing the decompressed data if encoded I/O fails. Benchmarks ========== I ran some benchmarks on send and receive of a zstd (level 3) compressed snapshot of a server's root filesystem which is about 23GB when compressed and 50GB when decompressed. Send v1: 0.41user 81.97system 2:21.71elapsed 58%CPU (0avgtext+0avgdata 2900maxresident)k 47182656inputs+0outputs (10major+119minor)pagefaults 0swaps Send compressed: 0.43user 60.53system 2:20.62elapsed 43%CPU (0avgtext+0avgdata 2836maxresident)k 47778864inputs+0outputs (8major+117minor)pagefaults 0swaps In this case, the bottleneck for send is reading the metadata trees and data from disk, so there's not much of a wall time improvement, but since the kernel doesn't have to decompress the data in the compressed case, it uses significantly less CPU and system time. Receive v1 into a filesystem with compress=none: 15.58user 62.36system 7:34.44elapsed 17%CPU (0avgtext+0avgdata 3028maxresident)k 104719648inputs+105333248outputs (1major+140minor)pagefaults 0swaps Receive v1 into a filesystem with compress-force=zstd: 15.45user 63.99system 5:11.57elapsed 25%CPU (0avgtext+0avgdata 3100maxresident)k 104587240inputs+105379328outputs (1major+143minor)pagefaults 0swaps Receive compressed into a filesystem with compress-force=zstd: 7.95user 44.53system 3:42.79elapsed 23%CPU (0avgtext+0avgdata 2992maxresident)k 46909600inputs+21603216outputs (2major+176minor)pagefaults 0swaps Without compressed receive, recompressing the data is still a wall time win because it requires much less I/O. However, compressed receive reduces the wall time even further. The v1 send stream is 50GB, and the v2 send stream is 23 GB. The v1 send stream compresses down to 17GB with zstd (level 3), so compressed send gets pretty close with no extra CPU overhead (the reason that compressed send is still larger is of course that we compress extents individually, which does not compress as efficiently as compressing the entire filesystem representation in one go). # ls -lh v1.send v1.send.zst compressed.send -rw-r--r-- 1 root root 23G Aug 17 12:34 compressed.send -rw-r--r-- 1 root root 50G Aug 17 12:13 v1.send -rw-r--r-- 1 root root 17G Aug 17 12:28 v1.send.zst Protocol Updates ================ This series makes some changes to the send stream protocol beyond adding the encoded write command/attributes and bumping the version. Namely, v1 has a 64k limit on the size of a write due to the 16-bit attribute length. This is not enough for encoded writes, as compressed extents may be up to 128k and cannot be split up. To address this, the BTRFS_SEND_A_DATA is treated specially in v2: its length is implicitly the remaining length of the command (which has a 32-bit length). This was the last bad of the options I considered. There are other commands that we've been wanting to add to the protocol: fallocate and FS_IOC_SETFLAGS. This series reserves their command and attribute numbers but does not implement kernel support for emitting them. However, it does implement support in receive for them, so the kernel can start emitting those whenever we get around to implementing them. Interface ========= For the send ioctl, stream version 2 is opt-in, and compressed writes are opt-in separately (but dependent on) stream version 2. Accordingly, `btrfs send` now accepts a `--stream-version` option and a `--compressed` option; the latter implies `--stream-version 2`. `btrfs receive` also accepts a `--force-decompress` option that forces the fallback to decompressing and writing the decompressed data. These options are provided to give the user flexibility in case they don't want their receiving filesytem to be compressed. Patches ======= The kernel patches are based on kdave/misc-next plus my "fs: interface for directly reading/writing compressed data" series. Patches 1-3 are improvements to the generic send code. Patches 4-7 do some preparation for stream v2 and compressed send. Patch 8 implements compressed send. Patch 9 modified the ioctl to accept the new flags and enable the new feature. Omar Sandoval (9): btrfs: send: get rid of i_size logic in send_write() btrfs: send: avoid copying file data btrfs: send: use btrfs_file_extent_end() in send_write_or_clone() btrfs: add send_stream_version attribute to sysfs btrfs: add send stream v2 definitions btrfs: send: write larger chunks when using stream v2 btrfs: send: allocate send buffer with alloc_page() and vmap() for v2 btrfs: send: send compressed extents with encoded writes btrfs: send: enable support for stream v2 and compressed writes fs/btrfs/ctree.h | 4 + fs/btrfs/inode.c | 6 +- fs/btrfs/send.c | 419 ++++++++++++++++++++++++++++--------- fs/btrfs/send.h | 33 ++- fs/btrfs/sysfs.c | 9 + include/uapi/linux/btrfs.h | 17 +- 6 files changed, 384 insertions(+), 104 deletions(-) The btrfs-progs patches were written by Boris Burkov. Patches 1-5 are preparation. Patch 6 implements encoded writes. Patch 7 implements the fallback to decompressing. Patch 8-9 implement the other commands. Patch 10 adds the new `btrfs send` options. Patch 11 adds a test case. Boris Burkov (11): btrfs-progs: receive: support v2 send stream larger tlv_len btrfs-progs: receive: dynamically allocate sctx->read_buf btrfs-progs: receive: support v2 send stream DATA tlv format btrfs-progs: receive: add send stream v2 cmds and attrs to send.h btrfs-progs: receive: add stub implementation for pwritev2 btrfs-progs: receive: process encoded_write commands btrfs-progs: receive: encoded_write fallback to explicit decode and write btrfs-progs: receive: process fallocate commands btrfs-progs: receive: process setflags ioctl commands btrfs-progs: send: stream v2 ioctl flags btrfs-progs: receive: add tests for basic encoded_write send/receive Makefile | 4 +- cmds/receive-dump.c | 31 +- cmds/receive.c | 402 +++++++++++++++++- cmds/send.c | 39 +- common/send-stream.c | 159 +++++-- common/send-stream.h | 7 + configure.ac | 1 + ioctl.h | 17 +- libbtrfsutil/btrfs.h | 17 +- send.h | 19 +- stubs.c | 24 ++ stubs.h | 50 +++ .../040-receive-write-encoded/test.sh | 114 +++++ 13 files changed, 832 insertions(+), 52 deletions(-) create mode 100644 stubs.c create mode 100644 stubs.h create mode 100755 tests/misc-tests/040-receive-write-encoded/test.sh Thanks!