On Thu, Oct 03, 2013 at 04:53:31PM +0200, Tanguy Bouzeloc wrote: > On 10/02/2013 11:19 AM, Ralf Baechle wrote: > >To my personal embarassment I have to admit that I knew about this since the > >day the syscall wrapper was written - but was considering it an acceptable > >bug ... > > > >Where it really bits is sigreturn and similar which use the following > >stunt: > > > > /* > > * Don't let your children do this ... > > */ > > __asm__ __volatile__( > > "move\t$29, %0\n\t" > > "j\tsyscall_exit" > > :/* no outputs */ > > :"r" (®s)); > > /* Unreached */ > > > >to keep the syscall return path from tampering with the return value. > > > >The scall*.S part of your patch is clearing TIF_NOERROR using a non-atomic > >LW/SW sequence. This needs to be done atomically or the thread's flags > >variable might get corrupted. This is complicated by MIPS I, R5900 and > >afair some older oddball not-quite MIPS II CPUs lacking LL/SC rsp. LLD/SCD. > > > > Ralf > > > > I discover the issue when changing the HZ of the kernel to 100HZ, in > this case the jiffies returned to the userland are the same as the > kernel ticks and it'll wrap after 5 minutes of uptime. With kernel > HZ at 250 or 1000H it'll make happen the ticks wrap after 230~260j. > > Unfortunately programs relying on ticks (they shouldn't but that > happens) have unpredictable behavior for 11.3s before the wrap. > > I can update the patch in order to access atomically the thread > flags, the point is ... it'll make the kernel incompatible with old > hardware. The price to pay: an ifdef ... #include <asm/sgidefs.H> ... #if (_MIPS_ISA > _MIPS_ISA_MIPS1) ... LL/SC code goes here #else ... LL/SC-less code goes here #endif But I'm wondering if we can move this hopefully relativly rare special case into do_notify_resume() and handle it in plain C there. Ralf
