在内核开发的过程中,经常会碰到内核崩溃,比如空指针异常,内存访问越界。通常我们只能靠崩溃之后打印出的异常调用栈信息来定位crash的位置和原因。总结下分析的方法和步骤。
通常oops发生之后,会在串口控制台或者dmesg日志输出看到如下的log,以某arm下linux内核的崩溃为例,
<2>[515753.310000] kernel BUG at net/core/skbuff.c:1846!<1>[515753.310000] Unable to handle kernel NULL pointer dereference at virtual address 00000000<1>[515753.320000] pgd = c0004000<1>[515753.320000] [00000000] *pgd=00000000<0>[515753.330000] Internal error: Oops: 817 [#1] PREEMPT SMP<0>[515753.330000] last sysfs file: /sys/class/net/eth0.2/speed<4>[515753.330000] module: http_timeout bf098000 4142...<4>[515753.330000] CPU: 0 Tainted: P (2.6.36 #2)<4>[515753.330000] PC is at __bug+0x20/0x28<4>[515753.330000] LR is at __bug+0x1c/0x28<4>[515753.330000] pc : [] lr : [ ] psr: 60000113<4>[515753.330000] sp : c0593e20 ip : c0593d70 fp : cf1b5ba0<4>[515753.330000] r10: 00000014 r9 : 4adec78d r8 : 00000006<4>[515753.330000] r7 : 00000000 r6 : 0000003a r5 : 0000003a r4 : 00000060<4>[515753.330000] r3 : 00000000 r2 : 00000204 r1 : 00000001 r0 : 0000003c<4>[515753.330000] Flags: nZCv IRQs on FIQs on Mode SVC_32 ISA ARM Segment kernel<4>[515753.330000] Control: 10c53c7d Table: 4fb5004a DAC: 00000017<0>[515753.330000] Process swapper (pid: 0, stack limit = 0xc0592270)<0>[515753.330000] Stack: (0xc0593e20 to 0xc0594000)<0>[515753.330000] 3e20: ce2ce900 c0543cf4 00000000 ceb4c400 000010cc c8f9b5d8 00000000 00000000<0>[515753.330000] 3e40: 00000001 cd469200 c8f9b5d8 00000000 ce2ce8bc 00000006 00000026 00000010...<4>[515753.330000] [ ] (PC is at __bug+0x20/0x28)<4>[515753.330000] [ ] (__bug+0x20/0x28) from [ ] (skb_checksum+0x3f8/0x400)<4>[515753.330000] [ ] (skb_checksum+0x3f8/0x400) from [ ] (et_isr+0x2b4/0x3dc [et])<4>[515753.330000] [ ] (et_isr+0x2b4/0x3dc [et]) from [ ] (et_txq_work+0x24/0x54 [et])<4>[515753.330000] [ ] (et_txq_work+0x24/0x54 [et]) from [ ] (et_tx_tasklet+0x14/0x298 [et])<4>[515753.330000] [ ] (et_tx_tasklet+0x14/0x298 [et]) from [ ] (tasklet_action+0x12c/0x174)<4>[515753.330000] [ ] (tasklet_action+0x12c/0x174) from [ ] (__do_softirq+0xfc/0x1a4)<4>[515753.330000] [ ] (__do_softirq+0xfc/0x1a4) from [ ] (irq_exit+0x60/0x64)<4>[515753.330000] [ ] (irq_exit+0x60/0x64) from [ ] (do_local_timer+0x60/0x74)<4>[515753.330000] [ ] (do_local_timer+0x60/0x74) from [ ] (__irq_svc+0x60/0x10c)<4>[515753.330000] Exception stack(0xc0593f68 to 0xc0593fb0)
在这里,我们着重关注下面几点:
Oops信息 kernel BUG at net/core/skbuff.c:1846! Unable to handle kernel NULL pointer dereference at virtual address 00000000
, 这里能够简要的告诉是什么问题触发了oops,如果是由代码直接调用BUG()/BUG_ON()一类的,还能给出源代码中触发的行号。
寄存器PC/LR的值 PC is at __bug+0x20/0x28 LR is at __bug+0x1c/0x28
, 这里PC是发送oops的指令, 可以通过LR找到函数的调用者
CPU编号和CPU寄存器的值 sp ip fp r0~r10
,
oops时,应用层的Process Process swapper (pid: 0, stack limit = 0xc0592270)
, 如果crash发生在内核调用上下文,这个可以用来定位对应的用户态进程
最重要的是调用栈,可以通过调用栈来分析错误位置
这里需要说明一点, skb_checksum+0x3f8/0x400
,在反汇编后,可以通过找到skb_checksum函数入口地址偏移0x3f8来精确定位执行点
在需要精确定位出错位置的时候,我们就需要用到反汇编工具objdump了。下面就是一个示例,
objdump -D -S xxx.o > xxx.txt
举个例子,比如我们需要寻找栈 (et_isr+0x2b4/0x3dc [et]) from [<bf11aa44>] (et_txq_work+0x24/0x54 [et])
,这里我们可以知道这个函数是在 [et] 这个obj文件中,那么我们可以直接去找 et.o ,然后反汇编 objdump -D -S et.o > et.txt
, 然后et.txt中就是反汇编后的指令。当然,单看汇编指令会非常让人头疼,我们需要反汇编指令和源码的一一对应才好分析问题。这就需要我们在编译compile的时候加上 -g 参数,把编译过程中的symbol和调试信息一并加入到最后obj文件中,这样objdump反汇编之后的文件中就包含嵌入的源码文件了。
对于内核编译来讲,就是需要在内核编译的根目录下,修改Makefile中 KBUILD_CFLAGS , 加上 -g 编译选项。
KBUILD_CFLAGS := -g -Wall -Wundef -Wstrict-prototypes -Wno-trigraphs \ -fno-strict-aliasing -fno-common \ -Werror-implicit-function-declaration \ -Wno-format-security \ -fno-delete-null-pointer-checks -Wno-implicit-function-declaration \ -Wno-unused-but-set-variable \ -Wno-unused-local-typedefs
下面是一份反编译完成后的文件的部分截取。我们可以看到,这里0x1f0是<et_isr>
这个函数的入口entry,c的源代码是在前面,后面跟的汇编代码是对应的反汇编指令
f0:et_isr(int irq, void *dev_id)#elsestatic irqreturn_t BCMFASTPATHet_isr(int irq, void *dev_id, struct pt_regs *ptregs)#endif{f0: e92d40f8 push {r3, r4, r5, r6, r7, lr}f4: e1a04001 mov r4, r1 struct chops *chops; void *ch; uint events = 0; et = (et_info_t *)dev_id; chops = et->etc->chops;f8: e5913000 ldr r3, [r1] ch = et->etc->ch; /* guard against shared interrupts */ if (!et->etc->up)fc: e5d32028 ldrb r2, [r3, #40] ; 0x28 struct chops *chops; void *ch; uint events = 0; et = (et_info_t *)dev_id; chops = et->etc->chops;: e5936078 ldr r6, [r3, #120] ; 0x78 ch = et->etc->ch;: e593507c ldr r5, [r3, #124] ; 0x7c /* guard against shared interrupts */ if (!et->etc->up): e3520000 cmp r2, #0c: 1a000001 bne 218 : e1a00002 mov r0, r2: e8bd80f8 pop {r3, r4, r5, r6, r7, pc} goto done; /* get interrupt condition bits */ events = (*chops->getintrevents)(ch, TRUE);: e5963028 ldr r3, [r6, #40] ; 0x28c: e1a00005 mov r0, r5: e3a01001 mov r1, #1: e12fff33 blx r3: e1a07000 mov r7, r0 /* not for us */ if (!(events & INTR_NEW))c: e2100010 ands r0, r0, #16: 08bd80f8 popeq {r3, r4, r5, r6, r7, pc} ET_TRACE(("et%d: et_isr: events 0x%x\n", et->etc->unit, events)); ET_LOG("et%d: et_isr: events 0x%x", et->etc->unit, events); /* disable interrupts */ (*chops->intrsoff)(ch);: e5963038 ldr r3, [r6, #56] ; 0x38: e1a00005 mov r0, r5c: e12fff33 blx r3 (*chops->intrson)(ch); }
在objdump反汇编出指令之后,我们可以根据调用栈上的入口偏移来找到对应的精确调用点。例如, (et_isr+0x2b4/0x3dc [et]) from [<bf11aa44>] (et_txq_work+0x24/0x54 [et])
, 我们可以知道调用点在et_isr
入口位置+0x2b4偏移 ,而刚才我们看到 et_isr
的入口位置是0x1f0 ,那就是说在 0x1f0+0x2b4=0x4a4
偏移位置。我们来看看,如下指令 4a4: e585007c str r0, [r5, #124] ; 0x7c
,其对应的源代码就是上面那一段c代码, skb->csum = skb_checksum(skb, thoff, skb->len - thoff, 0);
。而我们也知道,下一个调用函数的确是 skb_checksum
, 说明精确的调用指令是准确的。
ASSERT((prot == IP_PROT_TCP) || (prot == IP_PROT_UDP)); check = (uint16 *)(th + ((prot == IP_PROT_UDP) ?c: e3580011 cmp r8, #17: 13a0a010 movne sl, #16: 03a0a006 moveq sl, #6 offsetof(struct udphdr, check) : offsetof(struct tcphdr, check))); *check = 0;: e18720ba strh r2, [r7, sl] thoff = (th - skb->data); if (eth_type == HTON16(ETHER_TYPE_IP)) { struct iphdr *ih = ip_hdr(skb); prot = ih->protocol; ASSERT((prot == IP_PROT_TCP) || (prot == IP_PROT_UDP)); check = (uint16 *)(th + ((prot == IP_PROT_UDP) ?c: e087200a add r2, r7, sl: e58d2014 str r2, [sp, #20] offsetof(struct udphdr, check) : offsetof(struct tcphdr, check))); *check = 0; ET_TRACE(("et%d: skb_checksum: \n", et->etc->unit)); skb->csum = skb_checksum(skb, thoff, skb->len - thoff, 0);: e5952070 ldr r2, [r5, #112] ; 0x70: e58dc008 str ip, [sp, #8]c: e0612002 rsb r2, r1, r2a0: ebfffffe bl 0a4: e585007c str r0, [r5, #124] ; 0x7c *check = csum_tcpudp_magic(ih->saddr, ih->daddr,a8: e5953070 ldr r3, [r5, #112] ; 0x70static inline __wsumcsum_tcpudp_nofold(__be32 saddr, __be32 daddr, unsigned short len, unsigned short proto, __wsum sum){ __asm__(ac: e59dc008 ldr ip, [sp, #8]
有几点比较geek的地方需要注意:
函数调用栈的调用不一定准确(不知道why?可能因为调用过程是通过LR来反推到的,LR在执行过程中有可能被修改?),但是有一点可以确认,调用的点是准确的,也就是说调用函数不一定准,但是调用函数+偏移是能够找到准确的调入指令
inline的函数以及被优化的函数可能不会出现在调用栈上,在编译的时候因为优化的需要,会就地展开代码,这样就不会在这里有调用栈帧(stack frame)存在了