> On Fri, Nov 23, 2012 at 10:27:51AM +0400, Alexander Shiyan wrote:
> > > On Fri, Nov 23, 2012 at 07:40:30AM +0400, Alexander Shiyan wrote:
> > > [...]
> > > > > static unsigned long no_fiq_insn;
> > > > > +static int got_no_fiq_insn;
> > > > > @@ -78,11 +79,14 @@ void show_fiq_list(struct seq_file *p, int prec)
> > > > >
> > > > > void set_fiq_handler(void *start, unsigned int length)
> > > > > {
> > > > > -#if defined(CONFIG_CPU_USE_DOMAINS)
> > > > > - memcpy((void *)0xffff001c, start, length);
> > > > > -#else
> > > > > - memcpy(vectors_page + 0x1c, start, length);
> > > > > + unsigned long *addr = (void *)0xffff001c;
> > > > > +
> > > > > +#ifndef CONFIG_CPU_USE_DOMAINS
> > > > > + addr = vectors_page + 0x1c;
> > > > > #endif
> > > > > + if (!cmpxchg(&got_no_fiq_insn, 0, 1))
> > > > > + no_fiq_insn = *addr;
> > > > > + memcpy(addr, start, length);
> > > > > flush_icache_range(0xffff001c, 0xffff001c + length);
> > > > > if (!vectors_high())
> > > > > flush_icache_range(0x1c, 0x1c + length);
> > > > > @@ -126,8 +130,3 @@ EXPORT_SYMBOL(__set_fiq_regs); /* defined in
> > > > > fiqasm.S */
> > > > > -
> > > > > -void __init init_FIQ(void)
> > > > > -{
> > > > > - no_fiq_insn = *(unsigned long *)0xffff001c;
> > > >
> > > > it seems that this is wrong. In this case we have an uninitialized
> > > > variable and
> > > > sequential call claim_fiq and release_fiq could be fatal. FIXME please.
> > >
> > > Um... I don't think I understand, can you please elaborate?
> >
> > OK, I'll try to explain it.
> > At the end of the release_fiq function we have a call fiq_op. For the
> > default
> > handler - is a fiq_def_op function, and we call this function with the
> > option
> > "relinquish = 0", i.e. we want to restore old fiq_handler. But if we do not
> > call
> > set_fiq_handler never before, we will have an uninitialized no_fiq_insn
> > variable.
>
> It should not matter when or in what order anyone calls the
> set_fiq_handler(), since it stores "no FIQ instruction" into no_fiq_insn
> at its first invocation:
>
> if (!cmpxchg(&got_no_fiq_insn, 0, 1))
> no_fiq_insn = *addr;
>
> If we never called set_fiq_handler() before, during release_fiq() we'll:
>
> 1. Copy the initial instruction from the vector page to 'no_fiq_insn';
> 2. Copy the initial instruction from 'no_fiq_insn' to the vector page;
>
> So no_fiq_insn gets initialized, then we just instantly write the same
> value back.
got_no_fiq_insn also not initialized.
I think as a whole on this issue requires additional comments from Russell
as the author of implementation.
PS: In any case, I can reproduce it and check it out if you send me a patchset
to my email as an attachment.
---
