为了给即将入驻的 sourceforge 源腾地方，mirrors 将磁盘阵列的 XFS 扩充到 24T，现有 14T 可用空间。

删除无用内容

首先删除一些未同步成功又占据大量空间的源，释放了约 3T 磁盘空间。

  • android-src
  • android-releases
  • sourceforge
  • google-v8
  • chromiumos

扩容 XFS

XFS 号称能在线扩容，不过那是在 XFS 建立在 LVM 上时才可以。由于害怕 LVM 影响性能，mirrors 的 XFS 是直接建立在磁盘阵列的 GPT 上的。扩容 XFS 分为三步：

  1. 修改分区表
  2. 让内核重新载入分区信息，而这需要卸载文件系统
  3. 在线扩容 XFS

卸载磁盘阵列文件系统

以往卸载磁盘阵列的文件系统是全手工操作，折腾一次至少要十几分钟，其间磁盘阵列上的源都无法访问，整个网络服务（nginx 等）也会有一两分钟中断。这次我们使用了脚本，基本服务（nginx）每次中断时间不超过5秒，磁盘阵列的中断时间不超过2分钟（第一次中断了1分多钟是由于脚本里忘了 mount -a 重新挂载，收到报警短信后赶紧 mount -a）。

  1. 停止 LXC（因为 LXC 的根文件系统在磁盘阵列上）
  2. 杀掉所有服务进程
  3. 卸载已挂载的磁盘阵列文件系统（如果这一步出现问题，需要用 lsof 排查）
  4. 做想做的事
  5. 重新挂载文件系统
  6. 启动服务进程
  7. 启动 LXC

一开始的脚本是这样的：

sudo service rsync stop; sudo service vsftpd stop; sudo service nginx stop; mount | grep /dev/sdh1 | awk '{print $3}' | while read dir; do sudo umount $dir; sudo lsof $dir; done; sudo service nginx start; sudo service vsftpd start; sudo service rsync start

但发现 lsof 里还有很多 rsync 和 vsftpd 占着文件描述符，因此把 vsftpd 和 rsync 改成了杀气腾腾的 pkill：

sudo service nginx stop; sudo pkill vsftpd; sudo pkill rsync

还是有一些 rsync 进程发了 SIGTERM 信号仍无动于衷，因此改成了 SIGKILL（kill -9 所用的不可捕获杀进程信号）。

sudo service nginx stop; sudo pkill vsftpd; sudo pkill -SIGKILL rsync

现在 umount 成功了，但 rsync 启动失败了。这是由于没有删掉 rsyncd 的 pid 文件，强制删除即可。

sudo service nginx stop; sudo pkill vsftpd; sudo pkill -SIGKILL rsync; mount | grep /dev/sdh | awk '{print $3}' | while read dir; do sudo umount $dir; done; sudo service nginx start; sudo service vsftpd start; sudo rm -f /var/run/rsyncd.pid; sudo service rsync start

坑爹的 udev rule

扩容顺利完成了，不过 partprobe 之后，sdh 变成了跟 sdh1 一样的大小，分区表也不见了。重新启动 iscsi，问题依然如故。

$ sudo service nginx stop; sudo pkill vsftpd; sudo pkill -SIGKILL rsync; mount | grep /dev/sdh | awk '{print $3}' | while read dir; do sudo umount $dir; done; sudo /etc/init.d/open-iscsi restart; sudo service nginx start; sudo service vsftpd start; sudo rm -f /var/run/rsyncd.pid; sudo service rsync start
Stopping nginx: nginx.
Unmounting iscsi-backed filesystems: Unmounting all devices marked _netdev.
Disconnecting iSCSI targets:Logging out of session [sid: 1, target: iqn.2002-10.com.infortrend:raid.sn8223150.001, portal: 192.168.10.1,3260]
Logout of [sid: 1, target: iqn.2002-10.com.infortrend:raid.sn8223150.001, portal: 192.168.10.1,3260] successful.
.
Stopping iSCSI initiator service:.
Starting iSCSI initiator service: iscsid.
Setting up iSCSI targets:
Logging in to [iface: default, target: iqn.2002-10.com.infortrend:raid.sn8223150.001, portal: 192.168.10.1,3260] (multiple)
Login to [iface: default, target: iqn.2002-10.com.infortrend:raid.sn8223150.001, portal: 192.168.10.1,3260] successful.
.
Mounting network filesystems:.
Starting nginx: nginx.
Starting FTP server: vsftpd.
Starting rsync daemon: rsync.

$ sudo fdisk -l /dev/sdh

Disk /dev/sdh: 24189.3 GB, 24189254763008 bytes
255 heads, 63 sectors/track, 2940842 cylinders, total 47244638209 sectors
Units = sectors of 1 * 512 = 512 bytes
Sector size (logical/physical): 512 bytes / 512 bytes
I/O size (minimum/optimal): 512 bytes / 512 bytes
Disk identifier: 0x00000000

Disk /dev/sdh doesn't contain a valid partition table

$ sudo fdisk -l /dev/sdh1

Disk /dev/sdh1: 24189.3 GB, 24189254763008 bytes
255 heads, 63 sectors/track, 2940842 cylinders, total 47244638209 sectors
Units = sectors of 1 * 512 = 512 bytes
Sector size (logical/physical): 512 bytes / 512 bytes
I/O size (minimum/optimal): 512 bytes / 512 bytes
Disk identifier: 0x00000000

Disk /dev/sdh1 doesn't contain a valid partition table

线索在 syslog 里：

udevd[13846]: kernel-provided name 'sdh1' and NAME= 'sdh' disagree, please use SYMLINK+= or change the kernel to provide the proper name

原来是下面的 udev rule 捣的鬼，内核把 sdh1 和 sdh 都 probe 出来了，sdh1 占用了 sdh 这个名字（因为它们都符合 udev rules，不信可以自己执行 /lib/udev/scsi_id 验证）……

$ cat /etc/udev/rules.d/80-persistent-iscsi.rules
KERNEL=="sd*", SUBSYSTEM=="block", PROGRAM="/lib/udev/scsi_id --whitelisted --replace-whitespace /dev/$name", RESULT=="3600d0231000da93e75966be33fd9a2b4", NAME="sdh"

修改方法很简单，加个 %n 就行了。

$ cat /etc/udev/rules.d/80-persistent-iscsi.rules
KERNEL=="sd*", SUBSYSTEM=="block", PROGRAM="/lib/udev/scsi_id --whitelisted --replace-whitespace /dev/$name", RESULT=="3600d0231000da93e75966be33fd9a2b4", NAME="sdh%n"

扩容过程演示

下面我们演示将磁盘阵列 XFS 从 22T 扩容到 24T 的完整过程（为了文章清晰，删除了一些输出）。

1. 查看分区表信息。GPT 分区表要用 gdisk，而非 fdisk。

boj@mirrors:~$ sudo gdisk /dev/sdh
GPT fdisk (gdisk) version 0.8.5

Partition table scan:
MBR: protective
BSD: not present
APM: not present
GPT: present

Found valid GPT with protective MBR; using GPT.

Command (? for help): i
Using 1
Partition GUID code: EBD0A0A2-B9E5-4433-87C0-68B6B72699C7 (Microsoft basic data)
Partition unique GUID: 4592A37B-886C-40C9-A2AC-D9145B1D33D1
First sector: 2048 (at 1024.0 KiB)
Last sector: 47244640256 (at 22.0 TiB)
Partition size: 47244638209 sectors (22.0 TiB)
Attribute flags: 0000000000000000
Partition name: 'array'

2. 删除原有分区并新建分区、修改分区名称。注意分区起始扇区、GUID 都必须与原来的相同，否则文件系统无法识别！

Command (? for help): d 1
Using 1

Command (? for help): n
Partition number (1-128, default 1):
First sector (34-54684213214, default = 2048) or {+-}size{KMGTP}:
Last sector (2048-54684213214, default = 54684213214) or {+-}size{KMGTP}: 24T
Current type is 'Linux filesystem'
Hex code or GUID (L to show codes, Enter = 8300): EBD0A0A2-B9E5-4433-87C0-68B6B72699C7
Changed type of partition to 'Microsoft basic data'

Command (? for help): c 1
Using 1
Enter name: array

Command (? for help): i
Using 1
Partition GUID code: EBD0A0A2-B9E5-4433-87C0-68B6B72699C7 (Microsoft basic data)
Partition unique GUID: A5298F7D-A675-41DD-9AA3-BE119CA73DB3
First sector: 2048 (at 1024.0 KiB)
Last sector: 51539607552 (at 24.0 TiB)
Partition size: 51539605505 sectors (24.0 TiB)
Attribute flags: 0000000000000000
Partition name: 'array'

Command (? for help): w

Final checks complete. About to write GPT data. THIS WILL OVERWRITE EXISTING
PARTITIONS!!

Do you want to proceed? (Y/N): Y
OK; writing new GUID partition table (GPT) to /dev/sdh.
Warning: The kernel is still using the old partition table.
The new table will be used at the next reboot.
The operation has completed successfully.

3. 卸载磁盘阵列文件系统以便 partprobe 让内核重新载入分区表（详见上文）

boj@mirrors:~$ sudo lxc-list
RUNNING
  lxr
  pypi
  sync (auto)

boj@mirrors:~$ sudo lxc-stop -n pypi
boj@mirrors:~$ sudo lxc-stop -n sync
boj@mirrors:~$ sudo lxc-stop -n lxr

boj@mirrors:~$ sudo service nginx stop; sudo pkill vsftpd; sudo pkill -SIGKILL rsync; mount | grep /dev/sdh | awk '{print $3}' | while read dir; do sudo umount $dir; done; sudo partprobe; sudo mount -a; sudo service nginx start; sudo service vsftpd start; sudo rm -f /var/run/rsyncd.pid; sudo service rsync start
Stopping nginx: nginx.
Starting nginx: nginx.
Starting FTP server: vsftpd.
Starting rsync daemon: rsync.

boj@mirrors:~$ sudo lxc-start -n pypi -d
boj@mirrors:~$ sudo lxc-start -n sync -d
boj@mirrors:~$ sudo lxc-start -n lxr -d

4. 检查分区信息，使用 xfs_growfs 对 XFS 分区进行扩容。

boj@mirrors:~$ sudo gdisk -l /dev/sdh
GPT fdisk (gdisk) version 0.8.5

Partition table scan:
MBR: protective
BSD: not present
APM: not present
GPT: present

Found valid GPT with protective MBR; using GPT.
Disk /dev/sdh: 54684213248 sectors, 25.5 TiB
Logical sector size: 512 bytes
Disk identifier (GUID): 5E620981-EBFB-4EEA-BCF9-0052707F6859
Partition table holds up to 128 entries
First usable sector is 34, last usable sector is 54684213214
Partitions will be aligned on 2048-sector boundaries
Total free space is 3144607676 sectors (1.5 TiB)

Number Start (sector) End (sector) Size Code Name
1 2048 51539607552 24.0 TiB 0700 array

boj@mirrors:~$ sudo xfs_growfs /dev/sdh1
meta-data=/dev/sdh1 isize=256 agcount=22, agsize=268435455 blks
= sectsz=512 attr=2
data = bsize=4096 blocks=5905579776, imaxpct=5
= sunit=0 swidth=0 blks
naming =version 2 bsize=4096 ascii-ci=0
log =internal bsize=4096 blocks=521728, version=2
= sectsz=512 sunit=0 blks, lazy-count=1
realtime =none extsz=4096 blocks=0, rtextents=0
data blocks changed from 5905579776 to 6442450688

5. 检查扩容是否成功。

boj@mirrors:~$ df -lh /dev/sdh1
Filesystem Size Used Avail Use% Mounted on
/dev/sdh1 24T 11T 14T 43% /srv/ftp/ubuntu-old-releases

boj@mirrors:~$ sudo lxc-list
RUNNING
  lxr
  pypi
  sync (auto)

FROZEN

STOPPED
  mirror-lab
  mirror-lab_
  php-mirror
  root

最后提醒 mirrors 维护者，这类危险操作必须在 screen 或 byobu（screen 的封装，可以有选项卡）中进行，以免你的网络突然中断，脚本被 SIGHUP，将系统留在不可预测的状态。