2 The purpose of this document is to introduce the reader with vkernel debugging.
3 The vkernel architecture allows us to run DragonFly kernels in userland. These virtual
4 kernels can be paniced or otherwise abused, without affecting the host operating system.
6 To make things a bit more interesting, we will use a real life example.
9 ... I wrote a simple program that used the AIO interface. As it turned out we don't support
10 this feature, but at that point I didn't know.
12 $ gcc t_aio.c -o t_aio -Wall -ansi -pedantic
14 aio_read: Function not implemented
17 Ktrace'ing the process and seeing with my own eyes what was going on, seemed like a good idea.
18 Here comes the fun. I misread the ktrace(1) man page and typed:
24 (My intention was to track the system calls of t_aio, but what I typed would actually disable all traces from all process to the t_aio file.)
27 To setup a vkernel, please consult this [man page](http://leaf.dragonflybsd.org/cgi/web-man?command=vkernel§ion=ANY).
28 It's very straightforward.
30 # Reproduce the problem
31 We boot into our vkernel:
34 # ./boot/kernel -m 64m -r rootimg.01 -I auto:bridge0
38 And then try to reproduce the system freeze:
42 Fatal trap 12: page fault while in kernel mode
43 mp_lock = 00000001; cpuid = 1
44 fault virtual address = 0x0
45 fault code = supervisor read, page not present
46 instruction pointer = 0x1f:0x80aca52
47 stack pointer = 0x10:0x5709d914
48 frame pointer = 0x10:0x5709dbe0
49 processor eflags = interrupt enabled, resume, IOPL = 0
50 current process = 692 (ktrace)
51 current thread = pri 6
53 kernel: type 12 trap, code=4
55 CPU1 stopping CPUs: 0x00000001
57 Stopped at 0x80aca52: movl 0(%eax),%eax
60 This db> prompt is from ddb(4), the interactive kernel debugger.
63 fault virtual address = 0x0
65 field is indicative of a NULL pointer dereference inside the kernel.
67 Let's get a trace of what went wrong:
70 ktrdestroy(57082700,5709dc5c,0,57082700,5709dca0) at 0x80aca52
71 allproc_scan(80aca14,5709dc5c,be,2,0) at 0x80b2e91
72 sys_ktrace(5709dca0,6,0,0,57082700) at 0x80acffe
73 syscall2(5709dd40,6,57082700,0,0) at 0x8214b6d
74 user_trap(5709dd40,570940e8,8214185,0,8215462) at 0x8214d9c
75 go_user(5709dd38,0,0,7b,0) at 0x82151ac
78 sys_ktrace, allproc_scan, etc represent the function names. The hex values in parentheses are the first five items on the stack.
79 The last hex value is the instruction address. Since ddb doesn't really know how many arguments a function takes, it always prints five.
82 Quoting from vkernel(7):
84 It is possible to directly gdb the virtual kernel's process. It is recommended that you do a `handle SIGSEGV noprint' to ignore page faults processed by the virtual kernel itself and `handle SIGUSR1 noprint' to ignore signals used for simulating inter-processor interrupts (SMP build only).
86 You can add these two commands in your ~/.gdbinit to save yourself from typing them again and again.
89 handle SIGSEGV noprint
90 handle SIGUSR1 noprint
92 So we are going to attach to the vkernel process:
94 # ps aux | grep kernel
95 root 25408 0.0 2.3 1053376 17772 p0 IL+ 8:32PM 0:06.51 ./boot/kernel -m 64m -r rootimg.01 -I auto:bridge0
100 Let's get a trace from inside gdb:
103 #0 0x282d4c10 in sigsuspend () from /usr/lib/libc.so.6
104 #1 0x28287eb2 in sigsuspend () from /usr/lib/libthread_xu.so.2
105 #2 0x0821530a in stopsig (nada=24, info=0x40407d2c, ctxp=0x40407a4c) at /usr/src/sys/platform/vkernel/i386/exception.c:112
106 #3 <signal handler called>
107 #4 0x282d4690 in umtx_sleep () from /usr/lib/libc.so.6
108 #5 0x08213bde in cpu_idle () at /usr/src/sys/platform/vkernel/i386/cpu_regs.c:722
109 #6 0x00000000 in ?? ()
112 Why does it differ from the ddb's trace ?
113 Well, when the vkernel is sitting at a db> prompt all vkernel threads representing virtual cpu's except the one handling the db> prompt itself will be suspended in stopsig(). The backtrace only sees one of the N threads.
115 We need to do better this time. Let's break into the kernel _before_ it crashes. sys_ktrace() seems like a good candidate.
120 (gdb) break sys_ktrace
121 Breakpoint 1 at 0x80acf43: file ./machine/thread.h, line 83.
124 Next we type 'c' in the gdb prompt to resume vkernel execution:
129 Now we go to our vkernel and type the offending command:
133 Gdb stops the execution of vkernel and a message pops up in gdb buffer:
135 Breakpoint 1, sys_ktrace (uap=0x573e2ca0) at ./machine/thread.h:83
136 83 __asm ("movl %%fs:globaldata,%0" : "=r" (gd) : "m"(__mycpu__dummy));
139 We navigate through source code with the 'step' and 'next' gdb commands. They are identical, except that 'step' follows function calls. When we meet this call:
141 276 allproc_scan(ktrace_clear_callback, &info);
143 we 'step' inside it. alloproc_scan() iterates through the process list and applies the ktrace_clear_callback() to each one of them.
146 347 if (p->p_tracenode->kn_vp == info->tracenode->kn_vp) {
148 Here p is a pointer to the current process:
151 $1 = (struct proc *) 0x57098c00
153 Let's see if this process is traced:
155 (gdb) print p->p_tracenode
156 $2 = (struct ktrace_node *) 0x0
159 Oops. There is no trace to a vnode for this process. The code will try to access p->p_tracenode and is bound to crash. This is the zero virtual address we saw before.