On Sat, 2012-10-06 at 00:04 +0100, Al Viro wrote:
> On Fri, Oct 05, 2012 at 03:55:36PM +0100, James Bottomley wrote:
> > On Fri, 2012-10-05 at 15:48 +0100, Al Viro wrote:
> > > On Fri, Oct 05, 2012 at 02:44:24PM +0100, James Bottomley wrote:
> > > > On Fri, 2012-10-05 at 12:07 +0100, James Bottomley wrote:
> > > > > I tried out the code at
> > > > >
> > > > > git://git.kernel.org/pub/scm/linux/kernel/git/viro/signal.git
> > > > > experimental-kernel_thread
> > > > >
> > > > > and it gives me this panic on boot.
> > > >
> > > > OK, found the fix: the idle thread is a kernel thread, but it doesn't
> > > > come through kernel_thread(). The fix is to check for it (fortunately
> > > > it has the signal usp == 0).
> > >
> > > Um... I see, but I really wonder if that's the right fix. FWIW, sparc
> > > will have the same problem... Hell knows. OTOH, it's a nice way to
> > > get of implicit interplay between copy_thread() and idle_regs() - note
> > > that SMP architectures doing default idle_regs() need to be damn careful
> > > about what they do in their "is that kernel thread" logics; all-zeros
> > > pt_regs might give varying results on user_mode(regs) tests, etc.
> > > Might be better to go for
> > > if (p->flags & PF_KTHREAD) {
> > > if (!usp) {
> > > we are starting an idle thread
> > > } else {
> > > we are setting things up for kernel_thread()
> > > }
> > > } else {
> > > we are forking
> > > }
> > > kind of logics, looking at regs only in the last case. And to hell with
> > > (separate and overridable) idle_regs() once everything goes that way...
> >
> > But there's not a lot of point. forking an idle thread actually doesn't
> > care about any of the register execution setup because it never really
> > uses it to execute. That's why it was safe for us to use the user
> > thread setup ... I suppose the interior of the kernel thread case could
> > be conditioned on if (usp).
>
> BTW, speaking of parisc copy_thread()... Why the hell do we bother
> with *cregs = *pregs in userland case? It's a part of task_struct,
> after all, and we have copied that wholesale in arch_dup_task_struct().
>
> Another thing: why do we bother with
> STREG %r30,PT_GR21(%r1)
> in fork wrapper? We bloody well know what the offset will be, after all -
> right in the beginning of that sucker we'd done
> LDREG TI_TASK-THREAD_SZ_ALGN-FRAME_SIZE(%r30), %r1
> so we rely on %r30 having been (unsigned long)current_thread_info() +
> THREAD_SZ_ALGN + FRAME_SIZE. Then we add FRAME_SIZE again. IOW, the
> offset is a known constant. Hell, in child_return you rely on its
> value... While we are at it, I'm not sure you need to go through
> wrapper_exit on the way out in parent - saving cr27 can be done via
> e.g. r28 instead of r3, at which point you can simply branch to
> sys_clone() with no work left for wrapper_exit. *Child* obviously
> needs to restore these registers, so let it do that in child_return,
> but why bother in parent? After all, we are talking about the callee-saved
> registers, so sys_clone() is going to revert whatever changes it makes
> to them...
It sounds plausible. I'm checking your branch out now, modulo a couple
of compile failures:
arch/parisc/kernel/entry.S: Assembler messages:
arch/parisc/kernel/entry.S:1754: Error: Invalid operands
arch/parisc/kernel/process.c: In function 'copy_thread':
arch/parisc/kernel/process.c:256: error: 'FRAME_SIZE' undeclared (first use in this function)
arch/parisc/kernel/process.c:256: error: (Each undeclared identifier is reported only once
arch/parisc/kernel/process.c:256: error: for each function it appears in.)
arch/parisc/kernel/process.c:258: error: expected ')' before '{' token
arch/parisc/kernel/process.c:297: error: expected expression before '}' token
arch/parisc/kernel/process.c:252: warning: unused variable 'child_return'
arch/parisc/kernel/process.c:251: warning: unused variable 'ret_from_kernel_thread'
Fix up below.
> BTW, TIF_SYSCALL_TRACE and singlestepping are turned off in child, so I don't
> see any need for child_return to know where the parent had come from - it
> won't have anything to do in tracesys_exit anyway.
>
> I've folded your fixes and pushed the result; I've added (again, completely
> untested) optimizations along the lines of the above on top of those, as
> a separate commit. Comments?
Even with the patch applied, it's hanging on boot around the first
kthread spawns. I'm investigating.
James
---
diff --git a/arch/parisc/kernel/entry.S b/arch/parisc/kernel/entry.S
index f253f36..47fb6dd 100644
--- a/arch/parisc/kernel/entry.S
+++ b/arch/parisc/kernel/entry.S
@@ -1751,7 +1751,7 @@ ENTRY(sys_vfork_wrapper)
mfctl %cr27, %r28
STREG %r28, PT_CR27(%r1)
- b sys_vfork,%r2
+ b sys_vfork
copy %r1,%r26
ENDPROC(sys_vfork_wrapper)
diff --git a/arch/parisc/kernel/process.c b/arch/parisc/kernel/process.c
index 47eba15..61113c3 100644
--- a/arch/parisc/kernel/process.c
+++ b/arch/parisc/kernel/process.c
@@ -52,6 +52,7 @@
#include <asm/io.h>
#include <asm/asm-offsets.h>
+#include <asm/assembly.h>
#include <asm/pdc.h>
#include <asm/pdc_chassis.h>
#include <asm/pgalloc.h>
@@ -255,7 +256,7 @@ copy_thread(unsigned long clone_flags, unsigned long usp,
#endif
cregs->ksp = (unsigned long)stack + THREAD_SZ_ALGN + FRAME_SIZE;
- if (unlikely((p->flags & PF_KTHREAD)) {
+ if (unlikely(p->flags & PF_KTHREAD)) {
memset(cregs, 0, sizeof(struct pt_regs));
if (!usp) /* idle thread */
return 0;
--
To unsubscribe from this list: send the line "unsubscribe linux-parisc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
