Skip to contentSkip to conversation

Align VM locks, new spinlock patch

PATCH Juergen Doelle 1 participant 1 message Sep 26, 2001

Subsystems (6)
lore
Share & tools
lore
#0 lore raw
Quoted 3 lines · Andrew Morton

I think the coding can be simplified somewhat. How does this look?
...

Looks fine, it makes the code more common between UP and SMP. I consolidated the changes and appended the new version below.

The peak throughput results (MB/sec) for dbench on 2.4.10:

CPU 2.4.10 2.4.10 + improvement
spinlock patch by patch

 U  102,9   103,0          0,0%
 1   95,1    96,9          1,9%
 2  155,5   156,0          0,3%
 4  206,9   222,1          7,3%

8 194,9 233,7 19,9%

Throughput (MB/sec) for dbench with 8 clients on 8 CPUs:
CPU 2.4.10 2.4.10 + improvement
spinlock patch by patch
8 173,9 223,0 28,2%

Juergen

______________________________________________________________
Juergen Doelle
IBM Linux Technology Center - kernel performance
[email protected]

= = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =

include/linux/spinlock.h renamed from include/linux/spinlock.h.orig +16 -0
@@ -133,4 +133,20 @@
 extern int atomic_dec_and_lock(atomic_t *atomic, spinlock_t *lock);
 #endif
 
+#ifdef CONFIG_SMP
+#include <linux/cache.h>
+
+typedef union {
+    spinlock_t lock;
+    char fill_up[(SMP_CACHE_BYTES)];
+} spinlock_cacheline_t __attribute__ ((aligned(SMP_CACHE_BYTES)));
+
+#else	/* SMP */
+
+typedef struct {
+    spinlock_t lock;
+} spinlock_cacheline_t;
+
+
+#endif
 #endif /* __LINUX_SPINLOCK_H */
include/linux/swap.h renamed from include/linux/swap.h.orig +6 -2
@@ -86,7 +86,10 @@
 extern atomic_t nr_async_pages;
 extern atomic_t page_cache_size;
 extern atomic_t buffermem_pages;
-extern spinlock_t pagecache_lock;
+
+extern spinlock_cacheline_t pagecache_lock_cacheline;
+#define pagecache_lock (pagecache_lock_cacheline.lock)
+
 extern void __remove_inode_page(struct page *);
 
 /* Incomplete types for prototype declarations: */
@@ -159,7 +162,8 @@
 extern unsigned long swap_cache_find_success;
 #endif
 
-extern spinlock_t pagemap_lru_lock;
+extern spinlock_cacheline_t pagemap_lru_lock_cacheline;
+#define pagemap_lru_lock pagemap_lru_lock_cacheline.lock
 
 extern void FASTCALL(mark_page_accessed(struct page *));
 
include/asm-i386/smplock.h renamed from include/asm-i386/smplock.h.orig +2 -1
@@ -8,7 +8,8 @@
 #include <linux/sched.h>
 #include <asm/current.h>
 
-extern spinlock_t kernel_flag;
+extern spinlock_cacheline_t kernel_flag_cacheline;  
+#define kernel_flag kernel_flag_cacheline.lock      
 
 #define kernel_locked()		spin_is_locked(&kernel_flag)
 
arch/i386/kernel/i386_ksyms.c renamed from arch/i386/kernel/i386_ksyms.c.orig +1 -1
@@ -120,7 +120,7 @@
 
 #ifdef CONFIG_SMP
 EXPORT_SYMBOL(cpu_data);
-EXPORT_SYMBOL(kernel_flag);
+EXPORT_SYMBOL(kernel_flag_cacheline);
 EXPORT_SYMBOL(smp_num_cpus);
 EXPORT_SYMBOL(cpu_online_map);
 EXPORT_SYMBOL_NOVERS(__write_lock_failed);
arch/i386/kernel/smp.c renamed from arch/i386/kernel/smp.c.orig +1 -1
@@ -101,7 +101,7 @@
  */
 
 /* The 'big kernel lock' */
-spinlock_t kernel_flag = SPIN_LOCK_UNLOCKED;
+spinlock_cacheline_t kernel_flag_cacheline = {SPIN_LOCK_UNLOCKED};
 
 struct tlb_state cpu_tlbstate[NR_CPUS] = {[0 ... NR_CPUS-1] = { &init_mm, 0 }};
 
fs/buffer.c renamed from fs/buffer.c.orig +4 -1
@@ -81,7 +81,10 @@
 static rwlock_t hash_table_lock = RW_LOCK_UNLOCKED;
 
 static struct buffer_head *lru_list[NR_LIST];
-static spinlock_t lru_list_lock = SPIN_LOCK_UNLOCKED;
+
+static spinlock_cacheline_t lru_list_lock_cacheline = {SPIN_LOCK_UNLOCKED};
+#define lru_list_lock  lru_list_lock_cacheline.lock
+
 static int nr_buffers_type[NR_LIST];
 static unsigned long size_buffers_type[NR_LIST];
 
mm/filemap.c renamed from mm/filemap.c.orig +3 -2
@@ -46,12 +46,13 @@
 unsigned int page_hash_bits;
 struct page **page_hash_table;
 
-spinlock_t pagecache_lock ____cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
+spinlock_cacheline_t pagecache_lock_cacheline  = {SPIN_LOCK_UNLOCKED};
+
 /*
  * NOTE: to avoid deadlocking you must never acquire the pagecache_lock with
  *       the pagemap_lru_lock held.
  */
-spinlock_t pagemap_lru_lock ____cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
+spinlock_cacheline_t pagemap_lru_lock_cacheline = {SPIN_LOCK_UNLOCKED};
 
 #define CLUSTER_PAGES		(1 << page_cluster)
 #define CLUSTER_OFFSET(x)	(((x) >> page_cluster) << page_cluster)
mm/highmem.c renamed from mm/highmem.c.orig +2 -1
@@ -32,7 +32,8 @@
  */
 static int pkmap_count[LAST_PKMAP];
 static unsigned int last_pkmap_nr;
-static spinlock_t kmap_lock = SPIN_LOCK_UNLOCKED;
+static spinlock_cacheline_t kmap_lock_cacheline = {SPIN_LOCK_UNLOCKED};
+#define kmap_lock  kmap_lock_cacheline.lock
 
 pte_t * pkmap_page_table;