[PATCH 3/4] ppc32/kprobe: complete kprobe and migrate exception frame
Benjamin Herrenschmidt
benh at kernel.crashing.org
Tue Dec 13 10:19:47 EST 2011
On Mon, 2011-12-12 at 16:50 +0800, Tiejun Chen wrote:
> We can't emulate stwu since that may corrupt current exception stack.
> So we will have to do real store operation in the exception return code.
>
> Firstly we'll allocate a trampoline exception frame below the kprobed
> function stack and copy the current exception frame to the trampoline.
> Then we can do this real store operation to implement 'stwu', and reroute
> the trampoline frame to r1 to complete this exception migration.
>
> Signed-off-by: Tiejun Chen <tiejun.chen at windriver.com>
> ---
> arch/powerpc/kernel/entry_32.S | 26 ++++++++++++++++++++++++++
> 1 files changed, 26 insertions(+), 0 deletions(-)
>
> diff --git a/arch/powerpc/kernel/entry_32.S b/arch/powerpc/kernel/entry_32.S
> index 56212bc..d56e311 100644
> --- a/arch/powerpc/kernel/entry_32.S
> +++ b/arch/powerpc/kernel/entry_32.S
> @@ -1185,6 +1185,8 @@ recheck:
> bne- do_resched
> andi. r0,r9,_TIF_USER_WORK_MASK
> beq restore_user
> + andis. r0,r9,_TIF_DELAYED_KPROBE at h
> + bne- restore_kprobe
Same comment as earlier about name. Note that you're not hooking in the
right place. "recheck" is only reached if you -already- went out of the
normal exit path and only when going back to user space unless I'm
missing something (which is really the case you don't care about).
You need to hook into "resume_kernel" instead.
Also, we may want to simplify the whole thing, instead of checking user
vs. kernel first etc... we could instead have a single _TIF_WORK_MASK
which includes both the bits for user work and the new bit for kernel
work. With preempt, the kernel work bits would also include
_TIF_NEED_RESCHED.
Then you have in the common exit path, a single test for that, with a
fast path that skips everything and just goes to "restore" for both
kernel and user.
The only possible issue is the setting of dbcr0 for BookE and 44x and we
can keep that as a special case keyed of MSR_PR in the resume path under
ifdef BOOKE (we'll probably sanitize that later with some different
rework anyway).
So the exit path because something like:
ret_from_except:
.. hard disable interrupts (unchanged) ...
read TIF flags
andi with _TIF_WORK_MASK
nothing set -> restore
check PR
set -> do_work_user
no set -> do_work_kernel (kprobes & preempt)
(both loop until relevant _TIF flags are all clear)
restore:
#ifdef BOOKE & 44x test PR & do dbcr0 stuff if needed
... nornal restore ...
> do_user_signal: /* r10 contains MSR_KERNEL here */
> ori r10,r10,MSR_EE
> SYNC
> @@ -1202,6 +1204,30 @@ do_user_signal: /* r10 contains MSR_KERNEL here */
> REST_NVGPRS(r1)
> b recheck
>
> +restore_kprobe:
> + lwz r3,GPR1(r1)
> + subi r3,r3,INT_FRAME_SIZE; /* Allocate a trampoline exception frame */
> + mr r4,r1
> + bl copy_exc_stack /* Copy from the original to the trampoline */
> +
> + /* Do real stw operation to complete stwu */
> + mr r4,r1
> + addi r4,r4,INT_FRAME_SIZE /* Get kprobed entry */
> + lwz r5,GPR1(r1) /* Backup r1 */
> + stw r4,GPR1(r1) /* Now store that safely */
The above confuses me. Shouldn't you do instead something like
lwz r4,GPR1(r1)
subi r3,r4,INT_FRAME_SIZE
li r5,INT_FRAME_SIZE
bl memcpy
To start with, then you need to know the "old" r1 value which may or may
not be related to your current r1. The emulation code should stash it
into the int frame in an unused slot such as "orig_gpr3" (since that
only pertains to restarting syscalls which we aren't doing here).
Then you adjust your r1 and do something like
lwz r3,GPR1(r1)
lwz r0,ORIG_GPR3(r1)
stw r0,0(r3)
To perform the store, before doing the rest:
> + /* Reroute the trampoline frame to r1 */
> + subi r5,r5,INT_FRAME_SIZE
> + mr r1,r5
> +
> + /* Clear _TIF_DELAYED_KPROBE flag */
> + rlwinm r9,r1,0,0,(31-THREAD_SHIFT)
> + lwz r0,TI_FLAGS(r9)
> + rlwinm r0,r0,0,_TIF_DELAYED_KPROBE
> + stw r0,TI_FLAGS(r9)
> +
> + b restore
> +
> /*
> * We come here when we are at the end of handling an exception
> * that occurred at a place where taking an exception will lose
Cheers,
Ben.
More information about the Linuxppc-dev
mailing list