[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [Xen-devel] when timer go back in dom0 save and restore or migrate, PV domain hung

there's a clock_was_set called for each settimeofday. In latest kernel, clock_was_set will adjust CLOCK_REALTIME queue accordingly, while in 2.6.18 it's defined as a nop. That says, current domU would be unable to handle wallclock change, but newer kernel with pvops could.
for the issue reported in original thread, I agree that James should dig into the hang and explain the exact reason first.

From: Keir Fraser [mailto:keir.fraser@xxxxxxxxxxxxx]
Sent: Wednesday, November 26, 2008 10:58 PM
To: Tian, Kevin; 'James Song'; xen-devel@xxxxxxxxxxxxxxxxxxx
Subject: Re: [Xen-devel] when timer go back in dom0 save and restore or migrate, PV domain hung

So what happens if someone changes wallclock using 'date'? That's basically kind of what will appear to happen when s/r occurs.

 -- Keir

On 26/11/08 14:32, "Tian, Kevin" <kevin.tian@xxxxxxxxx> wrote:

hrtimer supports two timer bases: CLOCK_MONOTONIC and CLOCK_REALTIME. wall_to_monotonic is only added in former case, and for latter instead TOD is used directly per my reading. I did a quick search, and it looks that futex and ntp are using CLOCK_REALTIME. Also there's one vsyscall gate which can pass CLOCK_REALTIME from caller too.



From: Keir Fraser  [mailto:keir.fraser@xxxxxxxxxxxxx]
Sent: Wednesday, November 26,  2008 10:26 PM
To: Tian, Kevin; 'James Song';  xen-devel@xxxxxxxxxxxxxxxxxxx
Subject: Re: [Xen-devel] when timer go  back in dom0 save and restore or migrate, PV domain hung

hrtimers add wall_to_monotonic to xtime to get a  timesource that doesn't (or shouldn't!) warp.

 -- Keir

On  26/11/08 14:20, "Tian, Kevin" <kevin.tian@xxxxxxxxx>  wrote:

how about hrtimers? one mode is CLOCK_REALTIME, which uses  getnstimeofday as expiration. Once system time is changed either in local or  new machine, that expiration can't be adjusted. but i'm not sure whether it  still makes sense to try hrtimers in a guest.




From: Keir Fraser  [mailto:keir.fraser@xxxxxxxxxxxxx]  
Sent: Wednesday, November 26,  2008 10:11 PM
To:  Tian, Kevin; 'James Song';   xen-devel@xxxxxxxxxxxxxxxxxxx
Subject: Re: [Xen-devel]  when timer go  back in dom0 save and restore or migrate, PV domain  hung

The  problem hasn't been fully explained, but I can say  that PV guests  expect system time to jump across s/r and deal with that. For   example, Linux doesn't use Xen system time internally, but uses its  progress  to periodically update jiffies, which does not warp across  s/r.

We have  had problems corrupting wc_sec/wc_nsec in  xc_domain_restore.c, but that was  fixed some time  ago.

 -- Keir

On 26/11/08 14:00, "Tian,  Kevin"  <kevin.tian@xxxxxxxxx> wrote:

This is not a s/r or lm specific issue. For example, system  time  can be changed even when pv guest is running. Your patch only  hacks restore  point once, and wc_sec can still be changed later  when system time is  changed on-the-fly  again.

IIRC, pv guest can catch up wall clock change in timer  interrupt,  and time_resume will sync internal processed system  time with new system  time after restored. But I'm not sure whether  it's enough. Actually the more  interesting is the uptime  difference. For example, timer with expiration  calculated on  previous system time may wait nearly infinite if uptime among  two  boxes vary a lot. But I think such issue should have been considered   already, e.g. some user tool assistance. I think Keir can comment  better  here.

BTW, do you happen to know what exactly dom0 hangs on? In  some  busy loop to catch up time, or long delay to some critical  timer  expiration?




From:   xen-devel-bounces@xxxxxxxxxxxxxxxxxxx  [mailto:xen-devel-bounces@xxxxxxxxxxxxxxxxxxx]   On Behalf Of James  Song
Sent: Tuesday,  November 25,  2008 4:02 PM
To:    xen-devel@xxxxxxxxxxxxxxxxxxx
Subject:  [Xen-devel] when  timer go  back in dom0 save and restore or  migrate, PV domain  hung

   I   find PV domin hung, When we take those steps    
         1,  save PV  domain   
         2,   change system time of  PV domain back   
         3,  restore   a PV domain  
         1,  migrate  a PV domain  from Machine A to Machine   B
         2,  the system   time of Machine B is slower than Machine  A.
   the  problem is  wc_sec will be  change when system-time chanaged in dom0  or restore in a   slower-system-time machine, but when restoring, xen  don't  restore the wc_sec  of share_info from xenstore and use native   one. So guest os will hang.  
this patch will work for  this  issue.

 -- Song   Wei

diff -r  a5ed0dbc829f  tools/libxc/xc_domain_restore.c
---    a/tools/libxc/xc_domain_restore.c    Tue  Nov 18  14:34:14 2008  +0800
+++  b/tools/libxc/xc_domain_restore.c     Fri Nov 21   17:34:15 2008 +0800
@@ -328,6  +328,16   @@
     /* For info   only  */
     nr_pfns = 0;
+       //jsong@xxxxxxxxxx, james song
+      memset(&domctl, 0,   sizeof(domctl));
+     domctl.domain =   dom;
+     domctl.cmd    =    XEN_DOMCTL_restoredomain;
+    frc =   do_domctl(xc_handle,  &domctl);
+     if ( frc  != 0 )
+      {
+               ERROR("Unable   to set flag of  restore.");
+               goto   out;
+      }
     if  (   read_exact(io_fd, &p2m_size, sizeof(unsigned long))    )
@@ -1120,6 +1130,8    @@
     /* restore  saved  vcpu_info and arch  specific info   */
     MEMCPY_FIELD(new_shared_info,    old_shared_info, vcpu_info);
+       MEMCPY_FIELD(new_shared_info,  old_shared_info,   wc_nsec);
+     MEMCPY_FIELD(new_shared_info,    old_shared_info,   wc_sec);
      MEMCPY_FIELD(new_shared_info,   old_shared_info,    arch);
     /* clear  any  pending events and  the selector */
diff -r  a5ed0dbc829f  xen/arch/x86/time.c
---   a/xen/arch/x86/time.c     Tue Nov 18  14:34:14 2008 +0800
+++   b/xen/arch/x86/time.c     Fri Nov 21 17:34:15 2008  +0800
@@   -689,7 +689,6   @@
 void    update_vcpu_system_time(struct vcpu   *v)
      struct  cpu_time        *t;
@@ -703,7  +702,6   @@
     if (   u->tsc_timestamp ==  t->local_tsc_stamp   )
      u->tsc_timestamp       = t->local_tsc_stamp;
@@   -713,14  +711,19   @@
 void    update_domain_wallclock_time(struct domain    *d)
+      if(d->after_restore  )
+      {
+           d->after_restore  =  0;
+        goto   out;  //jsong@xxxxxxxxxx
+      }
      version_update_begin(&shared_info(d,    wc_version));
     shared_info(d,   wc_sec)  =  wc_sec +   d->time_offset_seconds;
     shared_info(d,    wc_nsec) =   wc_nsec;
      version_update_end(&shared_info(d,    wc_version));
@@   -751,7 +754,6  @@
     u64   x;
     u32 y,  _wc_sec,   _wc_nsec;
     struct domain    *d;
     x = (secs *  1000000000ULL)  + (u64)nsecs -   system_time_base;
     y  =  do_div(x,  1000000000);
@@ -1050,7 +1052,6   @@
 struct tm    wallclock_time(void)
     uint64_t    seconds;
     if (  !wc_sec    )
         return   (struct tm) { 0  };
diff -r a5ed0dbc829f   xen/common/domctl.c
---  a/xen/common/domctl.c      Tue Nov 18 14:34:14 2008 +0800
+++    b/xen/common/domctl.c    Fri Nov 21  17:34:15 2008  +0800
@@  -24,7 +24,6 @@
 #include   <asm/current.h>
 #include    <public/domctl.h>
 #include    <xsm/xsm.h>
 extern long    arch_do_domctl(
     struct  xen_domctl  *op,  XEN_GUEST_HANDLE(xen_domctl_t)  u_domctl);
@@  -315,6 +314,16    @@
         ret  =    0;
+     case XEN_DOMCTL_restoredomain:
+     {
+          struct domain   *d;
+         if ( (d  =   rcu_lock_domain_by_id(op->domain)) == NULL   )
+               break;
+          d->after_restore =    1;
+           rcu_unlock_domain(d);
+           break;
+     }
     case    XEN_DOMCTL_createdomain:
diff   -r a5ed0dbc829f  xen/include/public/domctl.h
---    a/xen/include/public/domctl.h    Tue Nov 18  14:34:14  2008  +0800
+++ b/xen/include/public/domctl.h      Fri Nov 21  17:34:15 2008 +0800
@@  -61,6 +61,7  @@
 #define  XEN_DOMCTL_destroydomain        2
 #define    XEN_DOMCTL_pausedomain           3
 #define   XEN_DOMCTL_unpausedomain        4
+#define   XEN_DOMCTL_restoredomain         51
 #define   XEN_DOMCTL_resumedomain         27
 #define    XEN_DOMCTL_getdomaininfo       5
diff -r   a5ed0dbc829f  xen/include/xen/sched.h
---   a/xen/include/xen/sched.h     Tue Nov 18 14:34:14 2008   +0800
+++  b/xen/include/xen/sched.h    Fri Nov 21  17:34:15   2008 +0800
@@ -231,6 +231,7   @@
      * cause a   deadlock.  Acquirers don't spin waiting; they    preempt.
      spinlock_t   hypercall_deadlock_mutex;
+    int  after_restore;    //jsong@xxxxxxxxxx
 struct    domain_setup_info
--Song    wei

Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.