On Sun, 29 Jul 2012, Eric Dumazet wrote:
> On Sun, 2012-07-29 at 12:10 +0200, Eric Dumazet wrote:
>
> > You can probably design something needing no more than 4 bytes per cpu,
> > and this thing could use non locked operations as bonus.
> >
> > like the following ...
>
> Coming back from my bike ride, here is a more polished version with
> proper synchronization/ barriers.
Hi Eric
I reworked your patch (it should be applied after my previous patch 3/3).
I replaced the rw-semaphore with a mutex. Instead of two pointers, I
changed it to one pointer and one bool variable.
I removed the barriers next to rcu (because, rcu works as a barrier) and
added a barrier when decrementing the percpu variable and when waiting for
percpu_count to be zero.
I tested performance of all implementation:
30.2s with no lock at all
32.2s with global rw-lock
30.6s with per-cpu rw-lock (my original implementation and Eric Dumazet's
implementation make no difference)
Mikulas
---
New percpu lock implementation
An alternative percpu lock implementation. The original idea by
Eric Dumazet <eric.dumazet@xxxxxxxxx>
The lock consists of an array of percpu unsigned integers, a boolean
variable and a mutex.
When we take the lock for read, we enter rcu read section, check for a
"locked" variable. If it is false, we increase a percpu counter on the
current cpu and exit the rcu section. If "locked" is true, we exit the
rcu section, take the mutex and drop it (this waits until a writer
finished) and retry.
Unlocking for read just decreases percpu variable. Note that we can
unlock on a difference cpu than where we locked, in this case the
counter underflows. The sum of all percpu counters represents the number
of processes that hold the lock for read.
When we need to lock for write, we take the mutex, set "locked" variable
to true and synchronize rcu. Since RCU has been synchronized, no
processes can create new read locks. We wait until the sum of percpu
counters is zero - when it is, there are no readers in the critical
section.
Signed-off-by: Mikulas Patocka <mpatocka@xxxxxxxxxx>
---
fs/block_dev.c | 15 ++----
include/linux/percpu-rwsem.h | 93 +++++++++++++++++++++----------------------
2 files changed, 53 insertions(+), 55 deletions(-)
Index: linux-3.5-fast/fs/block_dev.c
===================================================================
--- linux-3.5-fast.orig/fs/block_dev.c 2012-07-31 21:59:24.000000000 +0200
+++ linux-3.5-fast/fs/block_dev.c 2012-08-01 19:39:26.000000000 +0200
@@ -1602,13 +1602,12 @@ ssize_t blkdev_aio_read(struct kiocb *io
{
ssize_t ret;
struct block_device *bdev = I_BDEV(iocb->ki_filp->f_mapping->host);
- percpu_rwsem_ptr p;
- p = percpu_down_read(&bdev->bd_block_size_semaphore);
+ percpu_down_read(&bdev->bd_block_size_semaphore);
ret = generic_file_aio_read(iocb, iov, nr_segs, pos);
- percpu_up_read(&bdev->bd_block_size_semaphore, p);
+ percpu_up_read(&bdev->bd_block_size_semaphore);
return ret;
}
@@ -1627,11 +1626,10 @@ ssize_t blkdev_aio_write(struct kiocb *i
struct file *file = iocb->ki_filp;
struct block_device *bdev = I_BDEV(file->f_mapping->host);
ssize_t ret;
- percpu_rwsem_ptr p;
BUG_ON(iocb->ki_pos != pos);
- p = percpu_down_read(&bdev->bd_block_size_semaphore);
+ percpu_down_read(&bdev->bd_block_size_semaphore);
ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
if (ret > 0 || ret == -EIOCBQUEUED) {
@@ -1642,7 +1640,7 @@ ssize_t blkdev_aio_write(struct kiocb *i
ret = err;
}
- percpu_up_read(&bdev->bd_block_size_semaphore, p);
+ percpu_up_read(&bdev->bd_block_size_semaphore);
return ret;
}
@@ -1652,13 +1650,12 @@ int blkdev_mmap(struct file *file, struc
{
int ret;
struct block_device *bdev = I_BDEV(file->f_mapping->host);
- percpu_rwsem_ptr p;
- p = percpu_down_read(&bdev->bd_block_size_semaphore);
+ percpu_down_read(&bdev->bd_block_size_semaphore);
ret = generic_file_mmap(file, vma);
- percpu_up_read(&bdev->bd_block_size_semaphore, p);
+ percpu_up_read(&bdev->bd_block_size_semaphore);
return ret;
}
Index: linux-3.5-fast/include/linux/percpu-rwsem.h
===================================================================
--- linux-3.5-fast.orig/include/linux/percpu-rwsem.h 2012-07-31 21:47:25.000000000 +0200
+++ linux-3.5-fast/include/linux/percpu-rwsem.h 2012-08-01 19:32:53.000000000 +0200
@@ -3,75 +3,76 @@
#include <linux/rwsem.h>
#include <linux/percpu.h>
-
-#ifndef CONFIG_SMP
-
-#define percpu_rw_semaphore rw_semaphore
-#define percpu_rwsem_ptr int
-#define percpu_down_read(x) (down_read(x), 0)
-#define percpu_up_read(x, y) up_read(x)
-#define percpu_down_write down_write
-#define percpu_up_write up_write
-#define percpu_init_rwsem(x) (({init_rwsem(x);}), 0)
-#define percpu_free_rwsem(x) do { } while (0)
-
-#else
+#include <linux/rcupdate.h>
+#include <linux/delay.h>
struct percpu_rw_semaphore {
- struct rw_semaphore __percpu *s;
+ unsigned __percpu *counters;
+ bool locked;
+ struct mutex mtx;
};
-typedef struct rw_semaphore *percpu_rwsem_ptr;
-
-static inline percpu_rwsem_ptr percpu_down_read(struct percpu_rw_semaphore *sem)
+static inline void percpu_down_read(struct percpu_rw_semaphore *p)
{
- struct rw_semaphore *s = __this_cpu_ptr(sem->s);
- down_read(s);
- return s;
+retry:
+ rcu_read_lock();
+ if (unlikely(p->locked)) {
+ rcu_read_unlock();
+ mutex_lock(&p->mtx);
+ mutex_unlock(&p->mtx);
+ goto retry;
+ }
+ this_cpu_inc(*p->counters);
+ rcu_read_unlock();
}
-static inline void percpu_up_read(struct percpu_rw_semaphore *sem, percpu_rwsem_ptr s)
+static inline void percpu_up_read(struct percpu_rw_semaphore *p)
{
- up_read(s);
+ smp_wmb();
+ this_cpu_dec(*p->counters);
}
-static inline void percpu_down_write(struct percpu_rw_semaphore *sem)
+static inline unsigned int percpu_count(unsigned __percpu *counters)
{
+ unsigned total = 0;
int cpu;
- for_each_possible_cpu(cpu) {
- struct rw_semaphore *s = per_cpu_ptr(sem->s, cpu);
- down_write(s);
- }
+
+ for_each_possible_cpu(cpu)
+ total += ACCESS_ONCE(*per_cpu_ptr(counters, cpu));
+
+ return total;
}
-static inline void percpu_up_write(struct percpu_rw_semaphore *sem)
+static inline void percpu_down_write(struct percpu_rw_semaphore *p)
{
- int cpu;
- for_each_possible_cpu(cpu) {
- struct rw_semaphore *s = per_cpu_ptr(sem->s, cpu);
- up_write(s);
- }
+ mutex_lock(&p->mtx);
+ p->locked = true;
+ synchronize_rcu();
+ while (percpu_count(p->counters))
+ msleep(1);
+ smp_rmb(); /* paired with smp_wmb() in percpu_sem_up_read() */
}
-static inline int percpu_init_rwsem(struct percpu_rw_semaphore *sem)
+static inline void percpu_up_write(struct percpu_rw_semaphore *p)
{
- int cpu;
- sem->s = alloc_percpu(struct rw_semaphore);
- if (unlikely(!sem->s))
+ p->locked = false;
+ mutex_unlock(&p->mtx);
+}
+
+static inline int percpu_init_rwsem(struct percpu_rw_semaphore *p)
+{
+ p->counters = alloc_percpu(unsigned);
+ if (unlikely(!p->counters))
return -ENOMEM;
- for_each_possible_cpu(cpu) {
- struct rw_semaphore *s = per_cpu_ptr(sem->s, cpu);
- init_rwsem(s);
- }
+ p->locked = false;
+ mutex_init(&p->mtx);
return 0;
}
-static inline void percpu_free_rwsem(struct percpu_rw_semaphore *sem)
+static inline void percpu_free_rwsem(struct percpu_rw_semaphore *p)
{
- free_percpu(sem->s);
- sem->s = NULL; /* catch use after free bugs */
+ free_percpu(p->counters);
+ p->counters = NULL; /* catch use after free bugs */
}
#endif
-
-#endif
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
