n 06/25/2014 07:03 PM, Daniel Mack wrote:
On 06/25/2014 11:13 AM, AKASHI Takahiro wrote:
Current ioctl(CMD_MSG_RECV) contains an expensive cache operation. Even if
the data is quite small, it always executes D-Cache flush on a buffer
at minimum unit of one page in kdbus_conn_msg_install() /
kdbus_pool_slice_flush() .
In fact, flush_dcache_page() implementation is very expensive on ARM since
flushing is repeated at every cache line in a page. This may not be seen on
x86 because it doesn't have ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE on.
This patch replaces flush_dcache_page() by flush_kernel_vmap_range() with
the exact offset and size of slice. (Actually aligned to a cache-line size).
Signed-off-by: AKASHI Takahiro <[email protected]>
---
pool.c | 17 +++++++++++++++++
1 file changed, 17 insertions(+)
diff --git a/pool.c b/pool.c
index fcc54af..b723dd2 100644
--- a/pool.c
+++ b/pool.c
@@ -28,6 +28,13 @@
#include "pool.h"
#include "util.h"
+/*
+ * KDBUS_HACK3: Do dcache-flush only against the minimum range of slice
+ */
+#if KDBUS_HACK3
+#include <asm/cacheflush.h>
+#endif
+
/**
* struct kdbus_pool - the receiver's buffer
* @f: The backing shmem file
@@ -194,7 +201,13 @@ struct kdbus_pool_slice *kdbus_pool_slice_find(struct
kdbus_pool *pool,
int kdbus_pool_slice_alloc(struct kdbus_pool *pool,
struct kdbus_pool_slice **slice, size_t size)
{
+#if KDBUS_HACK3
+ /* FIXME: Assume that the cache line size is 128B. */
+#define DCACHE_ALIGN(size) ((size + 0x7f) & ~0x7fUL)
+ size_t slice_size = DCACHE_ALIGN(size);
+#else
size_t slice_size = KDBUS_ALIGN8(size);
+#endif
struct rb_node *n, *found = NULL;
struct kdbus_pool_slice *s;
int ret = 0;
@@ -620,6 +633,9 @@ exit_free:
*/
void kdbus_pool_slice_flush(const struct kdbus_pool_slice *slice)
{
+#if KDBUS_HACK3
+ flush_kernel_vmap_range((void *)slice->off, slice->size);
+#else
This seems to be ok. Later this week, I can test this on some old ARM
boards that reproducibly showed cache aliasing.
Thank you for your try.
Could you make this a real, non-hackish patch please that just removes
the old implementation and uses flush_kernel_vmap_range() instead,
without the KDBUS_HACK conditionals?
In fact, flush_kernel_vmap_range() is not common across all the architectures,
but a local function on some arch.
So I'm not sure I can create a generic fix for this hack at this moment.
-Takahiro AKASHI
Thanks,
Daniel
_______________________________________________
systemd-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/systemd-devel
