On Thu, 31 Aug 2017, Sebastian Andrzej Siewior wrote:

On 2017-08-31 21:11:08 [+0200], Thomas Gleixner wrote:

quoted

On Wed, 23 Aug 2017, Mike Galbraith wrote:

quoted

Use get_local_ptr() vs this_cpu_ptr().

Signed-off-by: Mike Galbraith <redacted>
---
 drivers/block/zram/zcomp.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

--- a/drivers/block/zram/zcomp.c
+++ b/drivers/block/zram/zcomp.c

@@ -120,7 +120,7 @@ struct zcomp_strm *zcomp_stream_get(stru
 {
 	struct zcomp_strm *zstrm;
 
-	zstrm = *this_cpu_ptr(comp->stream);
+	zstrm = *get_local_ptr(comp->stream);

This looks wrong. On mainline the calling code must have preemption disable
somehow, otherwise this_cpu_ptr() would not work.

This was introduced by Mike in a previous patch. The zstrm is only
accessed while the spinlock is held.

quoted

Looking at the call site it is;

	zram_slot_lock()
	  bit_spin_lock()

which is of course evading lockdep and everything else debugging wise.

Sebastian, do we have migration protection in bitlocked regions? And we
shpuld look into converting that into a spinlock on rt.

zram_lock_table() is bit_spin_lock() on !RT and spin_lock(&table->lock);
on RT. So this is done.
!RT has this running in a kmap_atomic() section so they have no
preemption there.
zcomp_stream_get() returns a per-CPU object which is protected with a
spinlock and only accessed locked.

So when we are inside a spinlocked section, why is this_cpu_ptr() not
working? That does not make sense.

Thanks,

	tglx