perf出现SIGBUS的coredump-编程知识

coredump信息

(gdb) bt full

#0 0x000055c37fa62c00 in perf_evsel__parse_sample (evsel=0x55c381223b00, event=0x7f144843ab30, data=0x7ffcbbcf6540) at util/evsel.c:1939

type = <optimized out>

swapped = <optimized out>

array = <optimized out>

max_size = <optimized out>

endp = <optimized out>

sz = <optimized out>

u = <optimized out>

#1 0x000055c37fa5c9ef in perf_evlist__parse_sample (evlist=<optimized out>, event=event@entry=0x7f144843ab30, sample=sample@entry=0x7ffcbbcf6540) at util/evlist.c:1415

evsel = <optimized out>

#2 0x000055c37fa9f169 in perf_session__deliver_event (session=0x55c38122f0a0, event=0x7f144843ab30, tool=0x55c37fd4f120 <record>, file_offset=3980080) at util/session.c:1483

sample = {ip = 0, pid = 2693265664, tid = 349153765, time = 0, addr = 94298151645472, id = 4242048, stream_id = 94298173534368, period = 139725088729728, weight = 140723459417504,

transaction = 140723459418880, insn_cnt = 94298148829675, cyc_cnt = 0, cpu = 571808, raw_size = 0, data_src = 0, phys_addr = 0, flags = 0, insn_len = 0, cpumode = 0 '\000', misc = 0,

insn = '\000' <repeats 15 times>, raw_data = 0x0, callchain = 0x0, branch_stack = 0x0, user_regs = {abi = 0, mask = 0, regs = 0x0, cache_regs = {0 <repeats 64 times>}, cache_mask = 0},

intr_regs = {abi = 0, mask = 0, regs = 0x0, cache_regs = {0 <repeats 45 times>, 1499604004643535104, 0, 0, 0, 0, 1914, 264, 0, 0, 0, 1499604004643535104, 0, 0, 0, 0, 94298149818197,

1914, 0, 0}, cache_mask = 0}, user_stack = {offset = 0, size = 0, data = 0x0}, read = {time_enabled = 94298175016936, time_running = 24048125519245, {group = {nr = 140723459418768,

values = 0x55c37faa2dd6 <queue_event+86>}, one = {value = 140723459418768, id = 94298148842966}}}}

ret = <optimized out>

#3 0x000055c37faa307d in do_flush (show_progress=true, oe=0x55c381235bb0) at util/ordered-events.c:244

head = 0x55c381235bd8

tmp = 0x55c381392378

iter = 0x55c3813923a0

limit = 18446744073709551615

prog = {title = 0x55c37fbaa388 "Processing time ordered events...", curr = 0, next = 119, step = 119, total = 1914, size = false}

ret = <optimized out>

last_ts = 24048125519245

head = <optimized out>

tmp = <optimized out>

iter = <optimized out>

limit = <optimized out>

last_ts = <optimized out>

prog = <optimized out>

ret = <optimized out>

__mptr = <optimized out>

__u = <optimized out>

__mptr = <optimized out>

#4 __ordered_events__flush (oe=0x55c381235bb0, how=OE_FLUSH__FINAL, timestamp=timestamp@entry=0) at util/ordered-events.c:323

err = <optimized out>

show_progress = true

str = {0x55c37fb897fc "NONE", 0x55c37fb90f55 "FINAL", 0x55c37fb8e667 "ROUND", 0x55c37fb90f5b "HALF ", 0x55c37fb90f61 "TOP ", 0x55c37fb90f67 "TIME "}

#5 0x000055c37faa3849 in __ordered_events__flush (timestamp=<optimized out>, how=<optimized out>, oe=<optimized out>) at util/ordered-events.c:339

err = <optimized out>

show_progress = <optimized out>

str = {0x55c37fb897fc "NONE", 0x55c37fb90f55 "FINAL", 0x55c37fb8e667 "ROUND", 0x55c37fb90f5b "HALF ", 0x55c37fb90f61 "TOP ", 0x55c37fb90f67 "TIME "}

err = <optimized out>

show_progress = <optimized out>

first = <optimized out>

last = <optimized out>

head = <optimized out>

__mptr = <optimized out>

__warned = 0

__ret_warn_once = <optimized out>

__ret_warn_on = <optimized out>

#6 ordered_events__flush (how=<optimized out>, oe=<optimized out>) at util/ordered-events.c:341

coredump分析

从coredump的栈来看，perf执行流程是正常的，搜了下SIGBUS可能的原因如下:

1、文件映射访问异常

这是 SIGBUS 在用户态最为常见的场景，也最容易触发，通常来说根本原因都是进程 mmap 了一个文件后，另外的进程把这个文件截断了，导致 mmap 出来的某些内存页超出文件的实际大小，访问那些超出的内存页就会触发 SIGBUS，具体来说有以下几种场景：
1、进程 mmap 一个文件后，其它进程 truncate 该文件到更小。
2、动态库更新，直接 cp 覆盖。
3、可执行文件更新，直接 cp 覆盖。

系统读取磁盘文件通常是按页映射到内存，出于效率考虑常常使用 copy on write 机制，所以文件映射之后，如果对应的文件 page 不存在了(truncated），也不见得会马上出问题，只有到访问时才会出错，因此有一定滞后期。

2、访问不对齐的内存##

X86 平台上访问不对齐的内存时，默认不会有问题，但用户可以手动设置 EFLAGS 把 CPU 设置为不允许非对齐的内存访问，此时如果出现不对齐的内存访问，SIGBUS 就会抛出，具体例子参看【3】。

3、Stack fault exception

这种场景非常罕见，通常是 OS 或者内存硬件问题，从 intel 的开发者文件来看，这种异常属于 trap，并不是我们用户态常说的 exception，这种异常有三种起因【4】：
1、canonical address violation.
Canonical address 指的是 64 位模式下，地址的高 48 ~ 64 不是全部是 0 或 1 的地址。
如果通过栈指针 rbp 或 rsp 访问了非 canonical address 内核就会发 stack fault trap，示例代码如下：