On Tue, Mar 22, 2005 at 01:01:45PM -0800, Andrew Morton wrote:
> I had the impression Richard planned on something more sophisticated
> than this??
Indeed. I suppose I can pass it along to avoid repeated "it doesn't build"
messages, but an smp kernel doesn't run. *All* processes spawned by init(1)
die with SIGILL very very shortly. I can only guess that there's been some
change to flushing (icache or tlb) that went wrong.
r~
===== include/asm-alpha/spinlock.h 1.14 vs edited =====
--- 1.14/include/asm-alpha/spinlock.h 2005-01-20 08:21:23 -08:00
+++ edited/include/asm-alpha/spinlock.h 2005-03-20 13:53:46 -08:00
@@ -15,7 +15,7 @@
*/
typedef struct {
- volatile unsigned int lock /*__attribute__((aligned(32))) */;
+ volatile unsigned int lock;
#ifdef CONFIG_DEBUG_SPINLOCK
int on_cpu;
int line_no;
@@ -23,40 +23,26 @@
struct task_struct * task;
const char *base_file;
#endif
-#ifdef CONFIG_PREEMPT
- unsigned int break_lock;
-#endif
} spinlock_t;
#ifdef CONFIG_DEBUG_SPINLOCK
-#define SPIN_LOCK_UNLOCKED (spinlock_t) {0, -1, 0, NULL, NULL, NULL}
-#define spin_lock_init(x) \
- ((x)->lock = 0, (x)->on_cpu = -1, (x)->previous = NULL, (x)->task = NULL)
+#define SPIN_LOCK_UNLOCKED (spinlock_t){ 0, -1, 0, NULL, NULL, NULL }
#else
-#define SPIN_LOCK_UNLOCKED (spinlock_t) { 0 }
-#define spin_lock_init(x) ((x)->lock = 0)
+#define SPIN_LOCK_UNLOCKED (spinlock_t){ 0 }
#endif
+#define spin_lock_init(x) do { *(x) = SPIN_LOCK_UNLOCKED; } while(0)
#define spin_is_locked(x) ((x)->lock != 0)
-#define spin_unlock_wait(x) ({ do { barrier(); } while ((x)->lock); })
-#define _raw_spin_lock_flags(lock, flags) _raw_spin_lock(lock)
+#define spin_unlock_wait(x) do { barrier(); } while ((x)->lock)
#ifdef CONFIG_DEBUG_SPINLOCK
extern void _raw_spin_unlock(spinlock_t * lock);
extern void debug_spin_lock(spinlock_t * lock, const char *, int);
extern int debug_spin_trylock(spinlock_t * lock, const char *, int);
-
-#define _raw_spin_lock(LOCK) debug_spin_lock(LOCK, __BASE_FILE__, __LINE__)
-#define _raw_spin_trylock(LOCK) debug_spin_trylock(LOCK, __BASE_FILE__, __LINE__)
-
-#define spin_lock_own(LOCK, LOCATION) \
-do { \
- if (!((LOCK)->lock && (LOCK)->on_cpu == smp_processor_id())) \
- printk("%s: called on %d from %p but lock %s on %d\n", \
- LOCATION, smp_processor_id(), \
- __builtin_return_address(0), \
- (LOCK)->lock ? "taken" : "freed", (LOCK)->on_cpu); \
-} while (0)
+#define _raw_spin_lock(LOCK) \
+ debug_spin_lock(LOCK, __BASE_FILE__, __LINE__)
+#define _raw_spin_trylock(LOCK) \
+ debug_spin_trylock(LOCK, __BASE_FILE__, __LINE__)
#else
static inline void _raw_spin_unlock(spinlock_t * lock)
{
@@ -68,19 +54,16 @@
{
long tmp;
- /* Use sub-sections to put the actual loop at the end
- of this object file's text section so as to perfect
- branch prediction. */
__asm__ __volatile__(
"1: ldl_l %0,%1\n"
- " blbs %0,2f\n"
- " or %0,1,%0\n"
+ " bne %0,2f\n"
+ " lda %0,1\n"
" stl_c %0,%1\n"
" beq %0,2f\n"
" mb\n"
".subsection 2\n"
"2: ldl %0,%1\n"
- " blbs %0,2b\n"
+ " bne %0,2b\n"
" br 1b\n"
".previous"
: "=&r" (tmp), "=m" (lock->lock)
@@ -91,22 +74,29 @@
{
return !test_and_set_bit(0, &lock->lock);
}
-
-#define spin_lock_own(LOCK, LOCATION) ((void)0)
#endif /* CONFIG_DEBUG_SPINLOCK */
+#define _raw_spin_lock_flags(lock, flags) _raw_spin_lock(lock)
+
/***********************************************************/
typedef struct {
- volatile unsigned int write_lock:1, read_counter:31;
-#ifdef CONFIG_PREEMPT
- unsigned int break_lock;
-#endif
-} /*__attribute__((aligned(32)))*/ rwlock_t;
+ volatile unsigned int lock;
+} rwlock_t;
+
+#define RW_LOCK_UNLOCKED (rwlock_t){ 0 }
-#define RW_LOCK_UNLOCKED (rwlock_t) { 0, 0 }
+#define rwlock_init(x) do { *(x) = RW_LOCK_UNLOCKED; } while(0)
-#define rwlock_init(x) do { *(x) = RW_LOCK_UNLOCKED; } while(0)
+static inline int read_can_lock(rwlock_t *lock)
+{
+ return (lock->lock & 1) == 0;
+}
+
+static inline int write_can_lock(rwlock_t *lock)
+{
+ return lock->lock == 0;
+}
#ifdef CONFIG_DEBUG_RWLOCK
extern void _raw_write_lock(rwlock_t * lock);
@@ -119,7 +109,7 @@
__asm__ __volatile__(
"1: ldl_l %1,%0\n"
" bne %1,6f\n"
- " or $31,1,%1\n"
+ " lda %1,1\n"
" stl_c %1,%0\n"
" beq %1,6f\n"
" mb\n"
@@ -142,7 +132,7 @@
" subl %1,2,%1\n"
" stl_c %1,%0\n"
" beq %1,6f\n"
- "4: mb\n"
+ " mb\n"
".subsection 2\n"
"6: ldl %1,%0\n"
" blbs %1,6b\n"
@@ -153,6 +143,28 @@
}
#endif /* CONFIG_DEBUG_RWLOCK */
+static inline int _raw_read_trylock(rwlock_t * lock)
+{
+ long regx;
+ int success;
+
+ __asm__ __volatile__(
+ "1: ldl_l %1,%0\n"
+ " lda %2,0\n"
+ " blbs %1,2f\n"
+ " subl %1,2,%2\n"
+ " stl_c %2,%0\n"
+ " beq %2,6f\n"
+ "2: mb\n"
+ ".subsection 2\n"
+ "6: br 1b\n"
+ ".previous"
+ : "=m" (*lock), "=&r" (regx), "=&r" (success)
+ : "m" (*lock) : "memory");
+
+ return success;
+}
+
static inline int _raw_write_trylock(rwlock_t * lock)
{
long regx;
@@ -162,10 +174,9 @@
"1: ldl_l %1,%0\n"
" lda %2,0\n"
" bne %1,2f\n"
- " or $31,1,%1\n"
- " stl_c %1,%0\n"
- " beq %1,6f\n"
" lda %2,1\n"
+ " stl_c %2,%0\n"
+ " beq %2,6f\n"
"2: mb\n"
".subsection 2\n"
"6: br 1b\n"
@@ -179,7 +190,7 @@
static inline void _raw_write_unlock(rwlock_t * lock)
{
mb();
- *(volatile int *)lock = 0;
+ lock->lock = 0;
}
static inline void _raw_read_unlock(rwlock_t * lock)
Richard Henderson wrote:
> On Tue, Mar 22, 2005 at 01:01:45PM -0800, Andrew Morton wrote:
>
>>I had the impression Richard planned on something more sophisticated
>>than this??
>
>
> Indeed. I suppose I can pass it along to avoid repeated "it doesn't build"
> messages, but an smp kernel doesn't run. *All* processes spawned by init(1)
> die with SIGILL very very shortly. I can only guess that there's been some
> change to flushing (icache or tlb) that went wrong.
I wonder if its 4-level page tables. I think DaveM, in another thread,
is chasing flush-<something> issues right now, on sparc64.
Jeff
On Wed, 23 Mar 2005 11:32:00 -0500
Jeff Garzik <[email protected]> wrote:
> I wonder if its 4-level page tables. I think DaveM, in another thread,
> is chasing flush-<something> issues right now, on sparc64.
It has to do with Hugh Dickin's patches which haven't been
integrated yet.
It may be the case that Alpha needs to override the pgd_addr_end(),
pmd_addr_end(), etc. macros (just like sparc64 and ia64), especially
if there are holes in the Alpha address space. But my recollection
is that there are not any.
Linus's tree is working just fine for me, it's just the new stuff
Hugh Dickins and Nick Piggin are working on.