From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mout.gmx.net ([212.227.17.21]:64114 "EHLO mout.gmx.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751168AbcG2MlI (ORCPT ); Fri, 29 Jul 2016 08:41:08 -0400 To: Filipe Manana , David Sterba , Chris Mason , Josef Bacik From: Qu Wenruo Subject: Btrfs send to send out metadata and data separately Cc: "linux-btrfs@vger.kernel.org" Message-ID: <07e7aea4-ebc7-1c47-34fb-daaae42ab245@gmx.com> Date: Fri, 29 Jul 2016 20:40:38 +0800 MIME-Version: 1.0 Content-Type: text/plain; charset=gbk; format=flowed Sender: linux-btrfs-owner@vger.kernel.org List-ID: Hi Filipe, and maintainers, I'm recently working on the root fix to free send from calling backref walk. My current idea is to send data and metadata separately, and only do clone detection inside the send subvolume. This method needs two new send commands: (And new send attribute, A_DATA_BYTENR) 1) SEND_C_DATA much like SEND_C_WRITE, with a little change in the 1st TLV. TLVs: A_DATA_BYTENR: bytenr of the data extent A_FILE_OFFSET: offset inside the data extent A_DATA: real data 2) SEND_C_CLONE_DATA A little like SEND_C_CLONE, with unneeded parameters striped TLVs: A_PATH: filename A_DATA_BYTENR: disk_bytenr of the EXTENT_DATA A_FILE_OFFSET: file offset A_FILE_OFFSET: offset inside the EXTENT_DATA A_CLONE_LEN: num_bytes of the EXTENT_DATA The send part is different in how to sending out a EXTENT_DATA. The send work follow is: 1) Found a EXTENT_DATA to send. Check rb_tree of "disk_bytenr". if "disk_bytenr" in rb_tree goto 2) Reflink data /* Initiate a SEND_C_DATA */ Send out the *whole* *uncompressed* extent of "disk_bytenr". Adds "disk_bytenr" into rb_tree 2) Reflink data /* Initiate a SEND_C_CLONE_DATA */ Filling disk_bytenr, offset and num_bytes, and send out the command. That's to say, send will send out extent data and referencer separately. So for kernel part, it's quite easy and *NO* time consuming backref walk ever. And no other part is modified. The main trick happens in the receive part. Receive will do the following thing first before recovering the subvolume/snapshot: 0) Create temporary dir for data extents Create a new dir with temporary name($data_extent), to put data extents into it. Then for SEND_C_DATA command: 1) Create file with file name $filename under $data_extent dir filename = $(printf "0x%x" $disk_bytenr) $disk_bytenr is the first u64 TLV of SEND_A_DATA command. 2) Write data into $data_extent/$filename Then handle the SEND_C_CLONE_DATA command It would be like xfs_io -f -c "reflink $data_extent/$disk_bytenr $extent_offset $file_offset $num_bytes" $filename disk_bytenr=2nd TLV (string converted to u64, with "0x%x") extent_offset=3rd TLV, u64 file_offset=4th TLV, u64 num_bytes=5th TLV, u64 filename=1th TLV, string Finally, after the snapshot/subvolume is recovered, remove the $data_extent directory. The whole idea is to completely remove the time consuming backref walk in send. So pros: 1) No backref walk, no soft lockup, no super long execution time Under worst case O(N^2), best case O(N) Memory usage worst case O(N), best case O(1) Where N is the number of reference to extents. 2) Almost the same metadata layout Including the overlap extents Cons: 1) Not full fs clone detection Such clone detection is only inside the send snapshot. For case that one extent is referred only once in the send snapshot, but also referred by source subvolume, then in the received subvolume, it will be a new extent, but not a clone. Only extent that is referred twice by send snapshot, that extent will be shared. (Although much better than disabling the whole clone detection) 2) Extra space usage Since it completely recovers the overlap extents 3) As many fragments as source subvolume 4) Possible slow recovery due to reflink speed. I am still concerned about the following problems: 1) Is it OK to add not only 1, but 2 new send commands? 2) Is such clone detection range change OK? Any ideas and suggestion is welcomed. Thanks, Qu