It turns out that our current __hash_page code will do a very hot busy-wait
loop waiting on _PAGE_BUSY to be cleared. It even does ldarx/stdcx in the
loop, which will bounce reservations around like crazy if there's more than
one CPU spinning on the same PTE (or even another PTE in the same
reservation granule). The end result is that each fault takes longer when
there's contention, which in turn increases the chance of another thread
hitting the same fault and also piling up. Not pretty.
There's two options here:
1. Do an out-of-line busy loop a'la spinlocks with just loads (no
reserves)
2. Just bail and refault if needed.
(2) makes sense here: If the PTE is busy, chances are it's in flux anyway
and the other code path making a change might just be ready to hash it.
This fixes a stampede seen on a large-ish system where a multithreaded
HPC app faults in the same text pages on several cpus at the same time.
Signed-off-by: Olof Johansson <olof@lixom.net>
Cc: Paul Mackerras <paulus@samba.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
bne- htab_wrong_access
/* Check if PTE is busy */
andi. r0,r31,_PAGE_BUSY
- bne- 1b
+ /* If so, just bail out and refault if needed. Someone else
+ * is changing this PTE anyway and might hash it.
+ */
+ bne- bail_ok
/* Prepare new PTE value (turn access RW into DIRTY, then
* add BUSY,HASHPTE and ACCESSED)
*/
/* Try all again */
b htab_insert_pte
+bail_ok:
+ li r3,0
+ b bail
+
htab_pte_insert_ok:
/* Insert slot number & secondary bit in PTE */
rldimi r30,r3,12,63-15