linux xfs的一次io异常导致的crash

8556395.925003] XFS (sdi): xfs_log_force: error - returned.
[8556407.425047] INFO: task umount: blocked for more than  seconds.----这个默认是120，该环境上是被人为设置1200
[8556407.425653] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[8556407.426050] umount          D ffff881fffdd47c0         0x00000080------------------------19286是pid，19283是ppid
[8556407.426056]  ffff883e39587d80  ffff883e39587fd8 ffff883e39587fd8
[8556407.426066]  ffff883e39587fd8 00000000000147c0 ffff881fd3260000 ffff881f860fec80
[8556407.426072]  ffff883ed79e1a20 ffff881f860fecc0 ffff881f860fece8 ffff881f860fec90------------这12个64位的值为栈里的值
[8556407.426078] Call Trace:
[8556407.426092]  [<ffffffff8163e879>] schedule+0x29/0x70
[8556407.426137]  [<ffffffffa07c6cd1>] xfs_ail_push_all_sync+0xc1/0x110 [xfs]
[8556407.426143]  [<ffffffff810a7610>] ? wake_up_atomic_t+0x30/0x30
[8556407.426165]  [<ffffffffa07b1b98>] xfs_unmountfs+0x68/0x190 [xfs]
[8556407.426186]  [<ffffffffa07b267b>] ? xfs_mru_cache_destroy+0x6b/0x90 [xfs]
[8556407.426204]  [<ffffffffa07b4a52>] xfs_fs_put_super+0x32/0x90 [xfs]
[8556407.426211]  [<ffffffff811e3026>] generic_shutdown_super+0x56/0xe0-------已经执行完fsnotify_umount_inodes
[8556407.426215]  [<ffffffff811e3477>] kill_block_super+0x27/0x70
[8556407.426218]  [<ffffffff811e37b9>] deactivate_locked_super+0x49/0x60
[8556407.426222]  [<ffffffff811e3db6>] deactivate_super+0x46/0x60
[8556407.426229]  [<ffffffff81200d25>] mntput_no_expire+0xc5/0x120
[8556407.426234]  [<ffffffff812025cb>] SyS_umount+0x15b/0x5b0
[8556407.426239]  [<ffffffff81649909>] system_call_fastpath+0x16/0x1b
[8556407.426242] sending NMI to all CPUs:

对应的函数为：

/*
 * Push out all items in the AIL immediately and wait until the AIL is empty.
 */
void
xfs_ail_push_all_sync(
    struct xfs_ail  *ailp)
{
    struct xfs_log_item    *lip;
    DEFINE_WAIT(wait);
 
    spin_lock(&ailp->xa_lock);
    while ((lip = xfs_ail_max(ailp)) != NULL) {
        prepare_to_wait(&ailp->xa_empty, &wait, TASK_UNINTERRUPTIBLE);
        ailp->xa_target = lip->li_lsn;
        wake_up_process(ailp->xa_task);
        spin_unlock(&ailp->xa_lock);
        schedule();
        spin_lock(&ailp->xa_lock);
    }
    spin_unlock(&ailp->xa_lock);
 
    finish_wait(&ailp->xa_empty, &wait);
}

排查的时候走了弯路，因为看的是：

crash> dis -l xfs_ail_push_all_sync+0xc1
/usr/src/debug/kernel-3.10.-327.22..el7/linux-3.10.-327.22..el7.x86_64/include/linux/spinlock.h:
0xffffffffa07c6cd1 <xfs_ail_push_all_sync+>: mov    %r13,%rdi

所以想当然地认为在第一个 获取spinlock的地方，

spin_lock(&ailp->xa_lock);---------以为在这hung住
    while ((lip = xfs_ail_max(ailp)) != NULL) {

然后查找这把锁：

  crash>  struct super_block.s_fs_info ffff883f7fe67800
  s_fs_info = 0xffff883f7fe67000
crash> mod -s xfs
     MODULE       NAME                   SIZE  OBJECT FILE
ffffffffa0803d40  xfs                    /lib/modules/3.10.-327.22..el7.x86_64/kernel/fs/xfs/xfs.ko
crash> struct -xo  xfs_mount.m_ail 0xffff883f7fe67000
struct xfs_mount {
  [ffff883f7fe67010] struct xfs_ail *m_ail;
}
 
 struct   xfs_mount.m_ail 0xffff883f7fe67000
  m_ail = 0xffff881f860fec80
 
crash> struct -xo  xfs_ail.xa_lock 0xffff881f860fec80
struct xfs_ail {
  [ffff881f860fecc0] spinlock_t xa_lock;
}

而且这把锁的值是：

crash> struct   xfs_ail.xa_lock 0xffff881f860fec80
  xa_lock = {
    {
      rlock = {
        raw_lock = {
          {
            head_tail = ,
            tickets = {
              head = ,
              tail =
            }
          }
        }
      }
    }
  }

我一看到这个是非0值，想当然认为是被持有了，为了验证我的想法，我甚至随便找了一把锁查看它的值，点背的时候，喝凉水都塞牙，我随手选择的锁是：

crash> p gc_lock
gc_lock = $ = {
  {
    rlock = {
      raw_lock = {
        {
          head_tail = ,
          tickets = {
            head = ,
            tail =
          }
        }
      }
    }
  }
}

后来仔细看了在没有开启 CONFIG_QUEUED_SPINLOCKS 的情况下，锁的定义才是如此，否则的话，是如下：

#ifndef __GENKSYMS__
typedef struct qspinlock {
    atomic_t    val;
} arch_spinlock_t;
#else
typedef u32 __ticketpair_t;
 
typedef struct arch_spinlock {
    union {
        __ticketpair_t    head_tail;
        struct __raw_tickets {
            __ticket_t head, tail;
        } tickets;
    };
} arch_spinlock_t;
#endif

所以要么0，要么1的情况只适合于第一种定义，以前因为业务默认配置的config问题，查的锁相关的crash，大多是第一种定义，但是当时没注意到这个细节，只是简单扫了一眼，浪费我两个小时，所以下盘不稳总会摔。

查看running状态的都没有持有这把锁，（其实按照第二种定义，是正常的，因为此时本就没人持有这把锁，因为head和tail相等）比较奇怪，后来找同事讨论，同事认为是while循环没跳出，所以获取锁的那行是schedule下面那行，还没有执行。

想起来自己曾经写过的博客《https://www.cnblogs.com/10087622blog/p/9558024.html》,应该是如此

void
xfs_ail_push_all_sync(
    struct xfs_ail  *ailp)
{
    struct xfs_log_item    *lip;
    DEFINE_WAIT(wait);
 
    spin_lock(&ailp->xa_lock);
    while ((lip = xfs_ail_max(ailp)) != NULL) {
        prepare_to_wait(&ailp->xa_empty, &wait, TASK_UNINTERRUPTIBLE);
        ailp->xa_target = lip->li_lsn;
        wake_up_process(ailp->xa_task);
        spin_unlock(&ailp->xa_lock);
        schedule();--------------------这个和while循环一直跳不出去，
        spin_lock(&ailp->xa_lock);
    }
    spin_unlock(&ailp->xa_lock);
 
    finish_wait(&ailp->xa_empty, &wait);
}

跳不出去的原因跟io相关，也就是wake_up 相关的内核线程之后，内核线程还是没有完成对应的任务：

crash> struct xfs_ail.xa_task 0xffff881f860fec80
  xa_task = 0xffff883f8901dc00
 
crash> task_struct.pid 0xffff883f8901dc00
  pid = 
 
crash> set
    PID:
COMMAND: "xfsaild/sdi"
   TASK: ffff883f8901dc00  [THREAD_INFO: ffff881f3d828000]
    CPU:
  STATE: TASK_INTERRUPTIBLE
crash> bt
PID:   TASK: ffff883f8901dc00  CPU:   COMMAND: "xfsaild/sdi"
 # [ffff881f3d82bcd0] __schedule at ffffffff8163df9b
 # [ffff881f3d82bd38] schedule at ffffffff8163e879
 # [ffff881f3d82bd48] schedule_timeout at ffffffff8163c295
 # [ffff881f3d82bdf8] xfsaild at ffffffffa07c6a4f [xfs]
 # [ffff881f3d82bec8] kthread at ffffffff810a661f
 # [ffff881f3d82bf50] ret_from_fork at ffffffff81649858

xfsaild 无法完成的原因也是因为IO异常:

附部分相关io异常的message日志：

kernel: sd :::: [sdi] CDB: Read()        4c b0
kernel: sd :::: [sdi] CDB: Read()         8a
kernel: blk_update_request: I/O error, dev sdi, sector
kernel: sd :::: rejecting I/O to offline device
kernel: XFS (sdi): metadata I/O error: block 0xee173de0 ("xfs_trans_read_buf_map") error  numblks
kernel: XFS (sdi): metadata I/O error: block 0x15da29d90 ("xfs_trans_read_buf_map") error  numblks
kernel: sd :::: rejecting I/O to offline device
kernel: XFS (sdi): xfs_imap_to_bp: xfs_trans_read_buf() returned error -.
kernel: XFS (sdi): metadata I/O error: block 0xeca8a480 ("xfs_trans_read_buf_map") error  numblks
kernel: sd :::: rejecting I/O to offline device
kernel: sd :::: rejecting I/O to offline device

对应的bug解决可以参照：https://bugzilla.redhat.com/show_bug.cgi?id=1267042