crash工具分析大型Linux服务器死锁实战

Linux服务器背景:
CPUS: 40
MEMORY: 127.6 GB
MACHINE: x86_64  (2199 Mhz)
Linux Kernel: 4.4.121
TASKS: 19411
其实这不算大型服务器,我见过大型的一般内存4T起步,300多个cpu.
故障背景:
客户发现系统卡死,手动触发了kdump.
使用下面命令发现有3092个D状态进程。
crash> ps | grep UN | wc -l
3092
小编头一回碰到这么多进程变成D状态,第一感觉应该就是死锁导致。
于是:
crash> ps | grep UN > UN  (导入UN文件慢慢玩)
发现UN里面有大量的ps、sudo这样的进程。
于是挑了一个ps进程查看它的堆栈:

crash> bt 50486

PID: 50486  TASK: ffff881171480c80  CPU: 18  COMMAND: "ps"

 #0 [ffff8810d1d9bce0] schedule at ffffffff815f353d

 #1 [ffff8810d1d9bd40] rwsem_down_read_failed at ffffffff815f61ea

 #2 [ffff8810d1d9bd98] call_rwsem_down_read_failed at ffffffff813271d4

 #3 [ffff8810d1d9bde0] down_read at ffffffff815f58b3

 #4 [ffff8810d1d9bde8] proc_pid_cmdline_read at ffffffff8126c5e8

 #5 [ffff8810d1d9be70] __vfs_read at ffffffff81202fd6

 #6 [ffff8810d1d9bee8] vfs_read at ffffffff8120360a

 #7 [ffff8810d1d9bf18] sys_read at ffffffff81204382


先看下proc_pid_cmdline_read 函数源码:

static ssize_t proc_pid_cmdline_read(struct file *file, char __user *buf,

                                     size_t _count, loff_t *pos)

{

        ...

        tsk = get_proc_task(file_inode(file));

        if (!tsk)

                return -ESRCH;

        mm = get_task_mm(tsk);

        down_read(&mm->mmap_sem);

        ...

}


void __sched down_read(struct rw_semaphore *sem)

从源码中知道 ps命令是在查询某个进程的状态时,执行down_read()没有获取到要查询进程struct mm_struct上的sem(读写锁)
所以下面开始汇编几个函数,先找到sem的值:

crash> dis proc_pid_cmdline_read

....

0xffffffff8126c5d7 <proc_pid_cmdline_read+135>: lea    0x68(%r15),%rax

0xffffffff8126c5db <proc_pid_cmdline_read+139>: mov    %rax,%rdi

0xffffffff8126c5de <proc_pid_cmdline_read+142>: mov    %rax,0x18(%rsp)

0xffffffff8126c5e3 <proc_pid_cmdline_read+147>: callq  0xffffffff815f58a0 


crash> dis down_read 0xffffffff815f58a0 : nopl 0x0(%rax,%rax,1) [FTRACE NOP] 0xffffffff815f58a5 <down_read+5>: mov %rdi,%rax 0xffffffff815f58a8 <down_read+8>: lock incq (%rax) 0xffffffff815f58ac <down_read+12>: jns 0xffffffff815f58b3 <down_read+19> 0xffffffff815f58ae <down_read+14>: callq 0xffffffff813271c0 0xffffffff815f58b3 <down_read+19>: retq
crash> dis call_rwsem_down_read_failed 0xffffffff813271c0 : push %rdi 0xffffffff813271c1 <call_rwsem_down_read_failed+1>: push %rsi 0xffffffff813271c2 <call_rwsem_down_read_failed+2>: push %rcx 0xffffffff813271c3 <call_rwsem_down_read_failed+3>: push %r8 0xffffffff813271c5 <call_rwsem_down_read_failed+5>: push %r9 0xffffffff813271c7 <call_rwsem_down_read_failed+7>: push %r10 0xffffffff813271c9 <call_rwsem_down_read_failed+9>: push %r11 0xffffffff813271cb <call_rwsem_down_read_failed+11>: push %rdx 0xffffffff813271cc <call_rwsem_down_read_failed+12>: mov %rax,%rdi 0xffffffff813271cf <call_rwsem_down_read_failed+15>: callq 0xffffffff815f6100

发现call_rwsem_down_read_failed() 中push %rdi ,%rdi就是 down_read()的参数,也就是sem的值。
下面把call_rwsem_down_read_failed()的堆栈读出来找参数:

crash> rd  ffff8810d1d9bd98 -e ffff8810d1d9bde0

ffff8810d1d9bd98:  ffffffff813271d4 ffff880000000000   .q2.............

ffff8810d1d9bda8:  0000000000024200 ffff88203f0226f0   .B.......&.? ...

ffff8810d1d9bdb8:  00000000004037ae 0000000000000078   .7@.....x.......

ffff8810d1d9bdc8:  0000000000000000 0000000000000020   ........ .......

ffff8810d1d9bdd8:  ffff88203c8fc0a8


发现 ffff88203c8fc0a8
就是sem的值,

crash> whatis -o mm_struct

...

   [100] spinlock_t page_table_lock;

   [104] struct rw_semaphore mmap_sem; (offset 0x68)


所以
mm_struct-> ffff88203c8fc040 (0xffff88203c8fc0a8 – 0x64)
所以马上就能知道ps命令在查询哪个进程时不能获取到sem锁。

crash> struct mm_struct.owner,mm_users  ffff88203c8fc040

  owner = 0xffff881f17b044c0

  mm_users = {

    counter = 3952


crash> struct task_struct.comm,pid 0xffff881f17b044c0 comm = "hmsserver\000st\000\000\000" pid = 21909

发现那个进程是21909,知道了是这个进程,有没有用呢,看天意了,先查询它的堆栈再说:

crash> bt 21909

PID: 21909  TASK: ffff881f17b044c0  CPU: 15  COMMAND: "hmsserver"

 #0 [ffff881f2ab43dd0] schedule at ffffffff815f353d

 #1 [ffff881f2ab43e30] rwsem_down_read_failed at ffffffff815f61ea

 #2 [ffff881f2ab43e80] call_rwsem_down_read_failed at ffffffff813271d4

 #3 [ffff881f2ab43ec8] down_read at ffffffff815f58b3

 #4 [ffff881f2ab43ed0] __do_page_fault at ffffffff81066d31

 #5 [ffff881f2ab43f28] do_page_fault at ffffffff81066dcb

 #6 [ffff881f2ab43f50] page_fault at ffffffff815fa822


发现并没有什么用,它自己都没能获取到sem锁.
这样就放弃吗? 生活还得继续,再说客户也不答应啊:)

从读写锁 rw_semaphore
开始柯南侦探模式:

crash> struct rw_semaphore -x ffff88203c8fc0a8

struct rw_semaphore {

  count = 0xffffffff00000001, 

  wait_list = {

    next = 0xffff880ee85bfe10, 

    prev = 0xffff88117aa67d50

  }, 

  ...

  owner = 0x0

 }


发现 owner
=
0x0
, 说明是一个读者持有了这把锁,这又增加了难度,

crash> struct -o rw_semaphore

struct rw_semaphore {

   [0] long count;

   [8] struct list_head wait_list;

  [24] raw_spinlock_t wait_lock;

  [28] struct optimistic_spin_queue osq;

  [32] struct task_struct *owner;

}


这样,然后我就把所有等待这个锁的进程全部找出来,把持有sem锁地址

ffff88203c8fc0a8( rw_sema
phore
)的进程也全部找出来,第一个输出结果有2947个进程,第二个的结果有3032个进程。一般持有这个sem锁的进程在3032-2947的进程上,我找了好几个小时都没看出端倪,好吧,下班吧,难搞哦。

crash> list rwsem_waiter.list -s rwsem_waiter.task,type -h 0xffff880ee85bfe10
crash> search -t  ffff88203c8fc0a8

从上面的分析来看,完全找不到线索,而且进程特别特别多,但是小编也没放弃,吃这碗饭,也没办法:)
于是看下sudo进程,找出一个sudo进程(sudo和ps看起来功能就完全没交叉):

PID: 22216  TASK: ffff8810ec908e80  CPU: 15  COMMAND: "sudo"

 #0 [ffff880f1d09fc78] schedule at ffffffff815f353d

 #1 [ffff880f1d09fcd8] schedule_preempt_disabled at ffffffff815f3e2e

 #2 [ffff880f1d09fce8] __mutex_lock_slowpath at ffffffff815f5665

 #3 [ffff880f1d09fd48] mutex_lock at ffffffff815f56f3

 #4 [ffff880f1d09fd58] rtnetlink_rcv at ffffffff81512135

 #5 [ffff880f1d09fd68] netlink_unicast at ffffffff815331c3

 #6 [ffff880f1d09fda0] netlink_sendmsg at ffffffff815335b3

 #7 [ffff880f1d09fe10] sock_sendmsg at ffffffff814e6cc6

 #8 [ffff880f1d09fe28] SYSC_sendto at ffffffff814e70f7

 #9 [ffff880f1d09ff50] entry_SYSCALL_64_fastpath at ffffffff815f7683


这里又弄出mutex_lock,真是一波未平,一波又起。
用同样的方法找mutex_waiter的地址

void __sched mutex_lock(struct mutex *lock)
crash> rd -SS ffff880f1d09fce8 -e ffff880f1d09fd48

ffff880f1d09fce8:  __mutex_lock_slowpath+149 ffff880126b63cf0 

ffff880f1d09fcf8:  ffff880fd53f7cf0 [ffff8810ec908e80:task_struct] 

ffff880f1d09fd08:  00000000024000c0 0000000000000180 

ffff880f1d09fd18:  rtnl_mutex       [ffff8810d2971800:kmalloc-256] 

ffff880f1d09fd28:  0000000000000014 0000000000000000 

ffff880f1d09fd38:  0000000000000000 [ffff8810a00ba000:kmalloc-2048]


从堆栈来看 ffff880fd53f7cf0
就是mutex_waiter, rtnl_mu
tex
就是mutex_lock的参数。

这样有些眉目了,在内核源码中找到 rtnl_mutex
是在下面文件中调用。

net/core/rtnetlink.c:


void rtnl_lock(void) { mutex_lock(&rtnl_mutex); }

所以看下mutex的owner是谁?

crash> struct mutex 0xffffffff81f0a180

struct mutex {

  count = {

    counter = -143

  }, 

...

  }, 

  wait_list = {

    next = 0xffff880f077d3dd0, 

    prev = 0xffff88020545fcf0

  }, 

  owner = 0xffff88103e4dc300, 

...

}


crash> struct task_struct.pid,comm 0xffff88103e4dc300 pid = 26299 comm = "kworker/7:2\000\000\000\000"

owner找到了,是kworker/7:2进程,看下它的堆栈,发现是i40e内核驱动模块持有了mutex锁。

crash> bt 26299

PID: 26299  TASK: ffff88103e4dc300  CPU: 7   COMMAND: "kworker/7:2"

 #0 [ffff880f4235fba8] schedule at ffffffff815f353d

 #1 [ffff880f4235fc08] schedule_timeout at ffffffff815f63e1

 #2 [ffff880f4235fcb0] msleep at ffffffff810ee599

 #3 [ffff880f4235fcc0] napi_disable at ffffffff814fd4be

 #4 [ffff880f4235fcd8] i40e_down at ffffffffa0628ab9 [i40e]

 #5 [ffff880f4235fd10] i40e_vsi_close at ffffffffa0628b73 [i40e]

 #6 [ffff880f4235fd30] i40e_close at ffffffffa0628ef1 [i40e]

 #7 [ffff880f4235fd38] i40e_pf_quiesce_all_vsi at ffffffffa0628c8a [i40e]

 #8 [ffff880f4235fd50] i40e_prep_for_reset at ffffffffa0628d81 [i40e]

 #9 [ffff880f4235fd70] i40e_do_reset at ffffffffa062ccf3 [i40e]

#10 [ffff880f4235fd90] i40e_service_task at ffffffffa062ead2 [i40e]

#11 [ffff880f4235fe38] process_one_work at ffffffff81097984

#12 [ffff880f4235fe78] worker_thread at ffffffff81098566

#13 [ffff880f4235fed0] kthread at ffffffff8109da59

#14 [ffff880f4235ff50] ret_from_fork at ffffffff815f7aaf


接下来查看内核i40e驱动代码:

i40e_service_task

  i40e_reset_subtask

    (rtnl_lock) i40e_do_reset

i40e_quiesce_vsi

  vsi->netdev->netdev_ops->ndo_stop(vsi->netdev);


发现在 i40e_reset_subtask
确实持有了锁.

list mutex_waiter.list -s mutex_waiter.task -h 0xffff880f077d3dd0 > mutex_list-0xffff880f077d3dd0

因为在上面发现 i40e_reset_subtask
函数持有了 rtnl_lock
(mutex锁),
所以看下这部分源码:

static void i40e_reset_subtask(struct i40e_pf *pf)

{

        u32 reset_flags = 0;


rtnl_lock(); ... if (reset_flags && !test_bit(__I40E_DOWN, &pf->state) && !test_bit(__I40E_CONFIG_BUSY, &pf->state)) i40e_do_reset(pf, reset_flags);
unlock: rtnl_unlock(); }

发现确实调用 i40e_do_reset
之后 schedule()切换走了。

void __rtnl_unlock(void)

{

        mutex_unlock(&rtnl_mutex);

}


void rtnl_unlock(void) { /* This fellow will unlock it for us. */ netdev_run_todo(); }

至此已经知道有问题的代码就在 i40e_reset_subtask
()函数中,
于是就从内核仓库中看看,这个bug有没有人已经修复了:
tig blame drivers/net/ethernet/intel/i40e/i40e_main.c
可以看到相关的commit ID 是373149f 或者dfc4ff6
分别看下:
git show 373149f


git show dfc4ff6


发现commit ID dfc4ff6已经修复了这个问题。
这样看起来分析完了,也解决了,但是有人会问,那些ps的进程与这个好像没有关系,现在看起来确实是,但是一定有个进程与ps堆栈中的sem锁有关,也与mutex锁有关,不过UN的进程实在太多,也没必要再分析了。

此公众号会分享更多实际解决案例。

如果有人对这个案例感兴趣,可以加我微信googuu,发30元红包(红包换知识),到时候我可以视频详细分享一下,一直有效.