swmore 发布的帖子

先拿swgdb给登上去看看在哪...

把-lm和-lm_slave放到两个.o后面试一试.
或者-Wl,--whole-archive -lm -lm_slave -Wl,--no-whole-archive试一试.

2018-12-21更新:

• 现在可以获取从核是否发生了SDLB Transform Exception或者DMA Descriptor Examination Warning了.
• 支持多次接收从核异常信号, 并拼接到原有的从核现场文件.
• 现在可以正确处理带有精确异常的从核错误, 并在错误处暂停主核程序.

现有从核调试用的小工具一件.
稳定版功能主要包括:

• 在SDLB Transform Exception和DMA Descriptor Examination Warning的情况下打印出从核的PC值.
• 在从核卡死的情况下打印出从核的PC值.
• 在Unknown Exception的情况下记录最后一个启动的从核函数的入口地址.
• 解析SDLB和DMA异常的现场.

测试版新增了:

• 在收到对应信号时打出所有从核所有通用寄存器的值.
• 可以唤醒通过halt指令(asm volatile("halt"))暂停的从核程序 (类似于断点?).

下载:

libspc.tgz
现在在psn上有发布目录了, 位于/home/export/online1/swmore/release/lib.

使用:

链接:

~在最终链接程序时添加-Wl,--whole-archive libspc.a -Wl,--no-whole-archive,-wrap,athread_init,-wrap,__real_athread_spawn,-wrap,__expt_handler (测试版为libspc-beta.a).~
现在没有测试版了, 目前看之前的测试版还算稳定, 我把输出格式改了一下.

source /home/export/online1/swmore/release/setenv #设置对应的环境变量
<原有的链接命令> $LINK_SPC #如果是在不能获取环境变量的编译脚本, 则echo $LINK_SPC看一眼里面是哪些选项贴进去也可以.

运行:

• 如果程序遇到从核/核组接口错, 将会打出错误现场信息.
• 如果程序收到SIGUSR1会打出错误现场信息.
• 如果链接了测试版程序收到SIGUSR2, 将会恢复halt的从核并打出错误现场信息.

发送SIGUSR1/SIGUSR2:

1. 可以通过bsignal -s 30 <作业号>向程序发送SIGUSR1, 通过bsignal -s 31 <作业号>向程序发送SIGUSR2.
2. 在使用swgdb等GDB类工具attach程序之后, 使用signal SIGUSR1/signal SIGUSR2向程序发送信号.

打出的报错信息:

样例屏幕输出:

Job <43262426> has been submitted to queue <q_sw_expr>
waiting for dispatch ...
dispatching ...
Before halt
Before halt
Before halt
Before halt
(NODE=vn000144, CG=0) wrote CPE spots to /home/export/online1/cesm06/dxh/workspace/cesm-scripts/spc/vn000144_cg0_cpe_spot.txt due to signal 31
(NODE=vn000144, CG=1) wrote CPE spots to /home/export/online1/cesm06/dxh/workspace/cesm-scripts/spc/vn000144_cg1_cpe_spot.txt due to signal 31
(NODE=vn000144, CG=2) wrote CPE spots to /home/export/online1/cesm06/dxh/workspace/cesm-scripts/spc/vn000144_cg2_cpe_spot.txt due to signal 31
(NODE=vn000144, CG=3) wrote CPE spots to /home/export/online1/cesm06/dxh/workspace/cesm-scripts/spc/vn000144_cg3_cpe_spot.txt due to signal 31
Waking up CPE 1!
After halt
Waking up CPE 1!
After halt
Waking up CPE 1!
Waking up CPE 1!
After halt
After halt

注意中间的4行类似于(NODE=vn000144, CG=0) wrote CPE spots to /home/export/online1/cesm06/dxh/workspace/cesm-scripts/spc/vn000144_cg0_cpe_spot.txt due to signal 31中包括的节点号(不是进程号!!!), 核组号, 错误现场输出文件路径, 输出原因 (30为SIGUSR1, 31为SIGUSR2, 55为从核报错).

样例报错文件:

==================DECODING ERR SPOT==================
Last function spawned 4ff04101b0
==================DECODE OF CPE PCs==================
  0:  4ff0410830   1:  4ff0410830   2:  4ff0410824   3:  4ff0410824   4:  4ff0410824   5:  4ff0410210   6:  4ff0410824   7:  4ff0410824 
  8:  4ff0410830   9:  4ff0410830  10:  4ff0410830  11:  4ff0410824  12:  4ff0410824  13:  4ff0410824  14:  4ff0410824  15:  4ff0410830 
 16:  4ff0410824  17:  4ff0410830  18:  4ff0410824  19:  4ff0410824  20:  4ff0410824  21:  4ff0410824  22:  4ff0410830  23:  4ff0410830 
 24:  4ff0410824  25:  4ff0410830  26:  4ff0410824  27:  4ff0410824  28:  4ff0410830  29:  4ff0410824  30:  4ff0410830  31:  4ff0410830 
 32:  4ff0410830  33:  4ff0410824  34:  4ff0410830  35:  4ff0410830  36:  4ff0410830  37:  4ff0410824  38:  4ff0410830  39:  4ff0410830 
 40:  4ff0410824  41:  4ff0410830  42:  4ff0410830  43:  4ff0410830  44:  4ff0410830  45:  4ff0410824  46:  4ff0410824  47:  4ff0410830 
 48:  4ff0410830  49:  4ff0410824  50:  4ff0410830  51:  4ff0410824  52:  4ff0410824  53:  4ff0410824  54:  4ff0410830  55:  4ff0410830 
 56:  4ff0410830  57:  4ff0410830  58:  4ff0410824  59:  4ff0410830  60:  4ff0410830  61:  4ff0410824  62:  4ff0410824  63:  4ff0410824 
#####################################################
## Can't determine the DMA and SDLB status due to  ##
## unknown reason, take the following info when job##
## shows corresponding exception.                  ##
#####################################################
==============DECODE OF SDLB ERROR SPOT==============
TC_SDLB_ERR_SPOT: 5800ff0000000080
REQ_TYPE: read
TC_SDLB_REQ_ADDR: 80
SRC_PE: 63
GRAIN: 8
SRC_TYPE: dma
OUT_OF_RANGE: no
OUT_OF_PERM: yes
===============DECODE OF DMACHK_FIELDS===============
DMACHK_FIELD0: 7000003
DMACHK_FIELD1: 6050
DMACHK_FIELD2: 200000
ERR_DETECTED: LDM unaligned
ERR_DETECTED: MEM unaligned
ERR_DETECTED: size unaligned
ERR_DETECTED: bsize unaligned
ERR_DETECTED: stride unaligned
SRC_PE: 5
TRASFER_SIZE: 3 (0x3)
BSIZE: 7 (0x7)
OP: DMA_GET
MODE: PE_MODE
REPLY_ADDR: 4
BCAST_MASK: 5
STRIDE: 6 (0x6)
MEM_ADDR: 1
LDM_ADDR: 2
DMA_PC[24:2]: 0x1040d5
=================DECODE OF CPE REGs================== #这一段及下面只有测试版才有.
CPE[00]:
   $0(  v0  ):
    hex           : 0000000000000000000000000000000000000000000000000000000000000000
    4B mark       :        7       6       5       4       3       2       1       0
    8B mark       :                3               2               1               0
    doublev4_value: 0.000000e+00, 0.000000e+00, 0.000000e+00, 0.000000e+00
    longv4_value  : 0, 0, 0, 0
    intv8_value   : 0, 0, 0, 0, 0, 0, 0, 0
   $1(  t0  ):
    hex           : 0000000000000000000000000000000000000000000000000000005000018d88
    4B mark       :        7       6       5       4       3       2       1       0
    8B mark       :                3               2               1               0
    doublev4_value: 1.697597e-312, 0.000000e+00, 0.000000e+00, 0.000000e+00
    longv4_value  : 343597485448, 0, 0, 0
    intv8_value   : 101768, 80, 0, 0, 0, 0, 0, 0
   $2(  t1  ):
    hex           : 000000000000000000000000000000000000000000000000000000500001c688
    4B mark       :        7       6       5       4       3       2       1       0
    8B mark       :                3               2               1               0
    doublev4_value: 1.697597e-312, 0.000000e+00, 0.000000e+00, 0.000000e+00
    longv4_value  : 343597500040, 0, 0, 0
    intv8_value   : 116360, 80, 0, 0, 0, 0, 0, 0
   $3(  t2  ):
    hex           : 000000000000000000000000000000000000000000000000000000500001c688
    4B mark       :        7       6       5       4       3       2       1       0
    8B mark       :                3               2               1               0
    doublev4_value: 1.697597e-312, 0.000000e+00, 0.000000e+00, 0.000000e+00
    longv4_value  : 343597500040, 0, 0, 0
#其实后面还有很多其他从核其他寄存器的数据.

解读报错信息:

Last function spawned:

最后启动spawn的一个从核函数, 任何情况下准确.

DECODE OF CPE PCs:

• 每行为8个从核当前的PC, 当发生有精确PC的异常时 (例如: LDM Access Exception, Unaligned Exception等等, 会打出the exact PC is...的异常), 由于从核函数被杀死 (从核回到slave_waiting_for_task函数), 不准确, 请参考这些Exception对应的打出的从核PC.
• 通过PC查询行号: addr2line -Cfe <应用程序的可执行文件> <PC值>.
• 通过PC看附近指令: sw5objdump -d <应用程序的可执行文件> | less, 使用/<PC值>+回车在得出的反汇编信息中搜索PC.

DECODE OF SDLB ERROR SPOT:

• 在发生SDLB Transform Exceptions时准确, 其他情况为随机值.
  • TC_SDLB_ERR_SPOT: SDLB错误现场原值, 仅供调试调试工具时使用.
  • REQ_TYPE: SDLB请求类型, 包括读/写/原子操作
  • TC_SDLB_REQ_ADDR: SDLB请求的地址
  • SRC_PE: 异常请求的源从核.
  • GRAIN: 请求的访存粒度(GLD/GST时精确, DMA时看不懂).
  • SRC_TYPE: 请求的来源(DMA请求为dma, 取指令为ibox, GLD/GST为cpe).
  • OUT_OF_RANGE: 访问是否越界.
  • OUT_OF_PERM: 访问是否越权.

DECODE OF DMACHK_FIELDS:

• 在发生DMA Descriptor Examination Warning时准确, 其他情况为随机值.
  • DMACHK_FIELDx: DMA错误现场原值, 仅供调试调试工具时使用.
  • ERR_DETECTED: DMA描述符异常的类型, 可能有多条.
  • SRC_PE: DMA请求的来源从核号.
  • TRASFER_SIZE, BSIZE, OP, MODE, REPLY_ADDR, BCAST_MASK, STRIDE, MEM_ADDR, LDM_ADDR: 对应于DMA/athread_get/put请求的各个参数.
  • DMA_PC[24:2]: DMA请求的PC地址的2到24位, 左移两位并补充高位地址后可以查处DMA报错的PC. 注意: athread_get/put时, DMA是在athread_get/put函数中发出的, 所以请参照DECODE OF CPE PCs一节来获取PC值, 它可能是在一个等待回答字的地方.

DECODE OF CPE REGs:

• 对从核寄存器中的值的解析, 适合调试汇编或者反汇编时对照.
• 结构为:

CPE[<从核号>]:
  $<寄存器号>(寄存器别名):
  hex: <通用寄存器的256位的16进制值, 从高到低排列>.
  4B/8B mark: 每4B/8B对应的长度有一个数, 便于从中截取int/long的16进制形式.
  doublev4_value: 寄存器作为doublev4时, 从低到高的4个double值(作为floatv4时, 该值仍可以参考).
  longv4_value: 寄存器作为4个long使用时,  从低到高的4个long值 (虽然本质上没有longv4类型).
  intv8_value: 寄存器作为intv8使用时, 从低到高的4个int值.

感谢:

• LAMMPS, GRAPES, CESM的诸位开发者写出来的bug提供了足够的的测试环境和素材.

@jixu 好像是明年6月份?

bsignal -s 11 <作业号>, 然后看打出来的PC...

-lm_slave

从核gdb...
不要奢望了...

ACC没有特别特别的了解...

8可能有点大.
看上去变成一次拷进32K了...
这样从核启动的延迟可能会变大一些...

所以都8月份了...
beacon呢?