From: Larry Woodman <lwoodman@redhat.com>
Date: Tue, 9 Dec 2008 11:11:07 -0500
Subject: [misc] ppc64: large sends fail with unix domain sockets
Message-id: 1228839067.23168.17.camel@dhcp-100-19-198.bos.redhat.com
O-Subject: [RHEL5-U4 patch] Fix large sends failure with unix domain sockets on PPC-64
Bugzilla: 461312
RH-Acked-by: Rik van Riel <riel@redhat.com>
RH-Acked-by: Pete Zaitcev <zaitcev@redhat.com>

On PPC64 SKB_MAX_ALLOC() ends up being 256KB due to the 64KB pagesize.

-----------------------------------------------------------------------
#define SKB_MAX_ORDER(X, ORDER) (((PAGE_SIZE << (ORDER)) - (X) - \
                                  sizeof(struct skb_shared_info)) & \
                                  ~(SMP_CACHE_BYTES - 1))

#define SKB_MAX_ALLOC           (SKB_MAX_ORDER(0, 2))
-----------------------------------------------------------------------

This can cause failures in unix domain sockets because that code uses
kmalloc() to allocate the buffers and the slabcache limits the largest
kmalloc pool to 128KB.

-------------------------------------------------------------------
if (size > SKB_MAX_ALLOC)
    size = SKB_MAX_ALLOC;

skb=sock_alloc_send_skb(sk,size,msg->msg_flags&MSG_DONTWAIT, &err);
-------------------------------------------------------------------
netperf: data send error: No buffer space available len was -1
-------------------------------------------------------------------

The fix for this problem is to simply increase the largest kmalloc()
pool to 256KB for PPC64 systems since they use a 64KB page size.

Fixes BZ461312

diff --git a/include/linux/kmalloc_sizes.h b/include/linux/kmalloc_sizes.h
index bda23e0..78a924e 100644
--- a/include/linux/kmalloc_sizes.h
+++ b/include/linux/kmalloc_sizes.h
@@ -19,7 +19,7 @@
 	CACHE(32768)
 	CACHE(65536)
 	CACHE(131072)
-#if (NR_CPUS > 512) || (MAX_NUMNODES > 256) || !defined(CONFIG_MMU)
+#if (NR_CPUS > 512) || (MAX_NUMNODES > 256) || defined(CONFIG_PPC_64K_PAGES) || !defined(CONFIG_MMU)
 	CACHE(262144)
 #endif
 #ifndef CONFIG_MMU