"David S. Miller" wrote:
>
> The soltuion you have chosen is not acceptable, the ordering is
> critical and is why all of this stuff is done this way in the first
> place. The members must be in _that_ order and the alignment must
> occur at that precise spot.
Yes, I know. But that patch did introduce a small difference from
the cache line POV: the first four entries of __tcp_listening_hash were
no longer in the same cache line as __tcp_ehash_size and friends.
> So please propose another patch, perhaps option #1 (the easiest) which
> preserves the existing structure layout.
This reorg should preserve things as you originally intended. I tend to
think that reorganising the structure is less fragile than fully
initialising the struct.
Still, if you prefer option 1 I have included that as well. Take your
pick. Both patches are tested and I have eyeballed the asm.
Question is: how many more of these are lurking...
Option 2
--------
--- linux-2.4.0-test10-pre3/include/net/tcp.h Sat Oct 14 17:02:04 2000
+++ linux-akpm/include/net/tcp.h Sun Oct 15 23:37:23 2000
@@ -90,6 +90,18 @@
};
extern struct tcp_hashinfo {
+ rwlock_t __tcp_lhash_lock;
+ atomic_t __tcp_lhash_users;
+ wait_queue_head_t __tcp_lhash_wait;
+ spinlock_t __tcp_portalloc_lock;
+
+ /*
+ * All the below members are written once at bootup and are
+ * never written again _or_ are predominantly read-access.
+ * Hence we align to a new cache line as all the preceding members
+ * are often dirty.
+ */
+
/* This is for sockets with full identity only. Sockets here will
* always be without wildcards and will have the following invariant:
*
@@ -97,8 +109,10 @@
*
* First half of the table is for sockets not in TIME_WAIT, second half
* is for TIME_WAIT sockets only.
+ *
*/
- struct tcp_ehash_bucket *__tcp_ehash;
+ struct tcp_ehash_bucket *__tcp_ehash
+ __attribute__((__aligned__(SMP_CACHE_BYTES)));
/* Ok, let's try this, I give up, we do need a local binding
* TCP hash as well as the others for fast bind/connect.
@@ -114,17 +128,6 @@
*/
struct sock *__tcp_listening_hash[TCP_LHTABLE_SIZE];
- /* All the above members are written once at bootup and
- * never written again _or_ are predominantly read-access.
- *
- * Now align to a new cache line as all the following members
- * are often dirty.
- */
- rwlock_t __tcp_lhash_lock
- __attribute__((__aligned__(SMP_CACHE_BYTES)));
- atomic_t __tcp_lhash_users;
- wait_queue_head_t __tcp_lhash_wait;
- spinlock_t __tcp_portalloc_lock;
} tcp_hashinfo;
#define tcp_ehash (tcp_hashinfo.__tcp_ehash)
Option 1
--------
--- linux-2.4.0-test10-pre3/net/ipv4/tcp_ipv4.c Sat Oct 14 17:02:04 2000
+++ linux-akpm/net/ipv4/tcp_ipv4.c Sun Oct 15 23:47:57 2000
@@ -75,7 +75,15 @@
void tcp_v4_send_check(struct sock *sk, struct tcphdr *th, int len,
struct sk_buff *skb);
+/*
+ * ALL members must be initialised to prevent gcc-2.7.2.3 miscompilation
+ */
struct tcp_hashinfo __cacheline_aligned tcp_hashinfo = {
+ __tcp_ehash: NULL,
+ __tcp_bhash: NULL,
+ __tcp_bhash_size: 0,
+ __tcp_ehash_size: 0,
+ __tcp_listening_hash: { NULL, },
__tcp_lhash_lock: RW_LOCK_UNLOCKED,
__tcp_lhash_users: ATOMIC_INIT(0),
__tcp_lhash_wait:
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [EMAIL PROTECTED]
Please read the FAQ at http://www.tux.org/lkml/
