12 Jan
2019
12 Jan
'19
5:13 p.m.
/ uses the system sleep call rather than the standard
C library
/ sleep (sleep (III)) because there is a critical race in the
/ C library implementation which could result in the process
/ pausing forever. This is a horrible bug in the UNIX process
/ control mechanism.
Quoted without comment from me!
Intriguing comment. I think your v6+ system probably has a lot of
PWB stuff in there. The libc source for sleep() in stock V6 is:
.globl _sleep
sleep = 35.
_sleep:
mov r5,-(sp)
mov sp,r5
mov 4(r5),r0
sys sleep
mov (sp)+,r5
rts pc
The PWB version uses something alarm/pause based, but apparently
gets it wrong:
.globl _sleep
alarm = 27.
pause = 29.
rti = 2
_sleep:
mov r5,-(sp)
mov sp,r5
sys signal; 14.; 1f
mov 4(r5),r0
sys alarm
sys pause
clr r0
sys alarm
mov (sp)+,r5
rts pc
1:
rti
I think the race occurs when an interrupt arrives between the sys alarm
and the sys pause lines, and the handler calls sleep again.
sleep() in the V7 libc is a much more elaborate C routine.
When I first read the race condition comment, I thought the issue would
be like that of write:
_write:
mov r5,-(sp)
mov sp,r5
mov 4(r5),r0
mov 6(r5),0f
mov 8(r5),0f+2
sys 0; 9f
bec 1f
jmp cerror
1:
mov (sp)+,r5
rts pc
.data
9:
sys write; 0:..; ..
This pattern appears in several V6 sys call routines, and would
not be safe when used in a context with signal based multi-
threading.