kern/111260: FreeBSD kernel dead lock and a solution

[Kris Kennaway]

The following reply was made to PR kern/111260; it has been noted by GNATS.

From: Kris Kennaway <kris at obsecurity.org>
To: Zhouyi Zhou <zhouzhouyi at ercist.iscas.ac.cn>
Cc: Kris Kennaway <kris at obsecurity.org>, freebsd-gnats-submit at FreeBSD.org
Subject: Re: kern/111260: FreeBSD kernel dead lock and a solution
Date: Sat, 7 Apr 2007 04:08:02 -0400

 On Sat, Apr 07, 2007 at 03:36:40PM +0800, Zhouyi Zhou wrote:
 > Dear Mr Kennaway
 >     It is sure to goto dead lock with simulatanouly tests running after several days, I use FreeBSD's DEBUG_LOCKS options
 > with lk_stack to infer where the thead get the lock, and when the thread is not swapped out, I use 
 > ((struct i386_frame *)(struct thread * (0xc*****))->td_pcb->pcb_ebp)->f_frame->f_frame-> ...... ->f_retaddr
 > is infer where lead the thread into sleep.
 >    Besides all above, to find the reason that lead to dead lock, I modified 
 > sys/stack.h to:
 > 32 #define STACK_MAX       50      
 > 33 
 >  34 struct sbuf;
 >  35 
 >  36 struct stack {
 >  37         int             depth;
 >  38         vm_offset_t     pcs[STACK_MAX];
 > 39         vm_offset_t     arg0[STACK_MAX];
 > 40 };
 > and the function stack_save in file i386/i386/db_trace.c
 > to save the first argument beside the return address.
 >    And In the case of tracing the swapped out thread, I modified the thread struct  in sys/proc.h and msleep function 
 > in kern/kern_synch.c to save the calling stack when the thread is going to sleep:
 > 241 struct thread {
 > 242         struct proc     *td_proc;       /* (*) Associated process. */
 > 243         struct ksegrp   *td_ksegrp;
 > .....
 > 327         struct stack  td_stack
 > 328  }
 > 
 > 118 int
 > 119 msleep(ident, mtx, priority, wmesg, timo)
 > 120         void *ident;
 > 121         struct mtx *mtx;
 > 122         int priority, timo;
 > 123         const char *wmesg;
 > 124 {
 > 125         struct thread *td;
 > 126         struct proc *p;
 > 127         int catch, rval, flags;
 > 128         WITNESS_SAVE_DECL(mtx);
 > 129 
 > 130         td = curthread;
 > 131         stack_save(td->td_stack);
 >      It is absolutely evidence that it is the 
 >  462 if (p->p_sysent->sv_copyout_strings)
 >  463 stack_base = (*p->p_sysent->sv_copyout_strings)(imgp); 
 > in do_execve that lead to dead lock.
 
 These are your conclusions, I am asking for the stack traces that lead
 you to them so that we can verify your observations.
 
 Kris