On Wed, 2011-03-30 at 14:27 +0100, Eric Dumazet wrote:

Le mercredi 30 mars 2011 à 14:17 +0100, Maxin John a écrit :

quoted

I have compiled the kernel with below given modification in .config

CONFIG_CMDLINE="uhash_entries=256"

After booting with the new kernel, the "kmemleak" no longer complains
about the "udp_table_init".
However it do report another possible leak :)

debian-mips:~# cat /sys/kernel/debug/kmemleak
unreferenced object 0x8f085000 (size 4096):
  comm "swapper", pid 1, jiffies 4294937670 (age 1043.280s)
  hex dump (first 32 bytes):
    00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
    00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
  backtrace:
    [<801ac7a8>] __kmalloc+0x130/0x180
    [<80532500>] flow_cache_cpu_prepare+0x50/0xa8
    [<8052378c>] flow_cache_init_global+0x90/0x138
    [<80100584>] do_one_initcall+0x174/0x1e0
    [<8050c348>] kernel_init+0xe4/0x174
    [<80103d4c>] kernel_thread_helper+0x10/0x18
debian-mips:~#

Hmm, then MIPS kmemleak port might have a problem with percpu data ?

fcp->hash_table = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));

fcp is a per cpu "struct flow_cache_percpu"

I can't figure out what it is. Kmemleak uses this block for scanning the
percpu data:

	for_each_possible_cpu(i)
		scan_block(__per_cpu_start + per_cpu_offset(i),
			   __per_cpu_end + per_cpu_offset(i), NULL, 1);

The __per_cpu_* symbols seem to be correctly defined in the MIPS
vmlinux.lds.S as it uses the PERCPU macro directly.

Other chunks allocated via pcpu_mem_alloc() should be tracked by
kmemleak and either reported as leaks or scanned (and not reporting
subsequent blocks referred from the percpu memory).

-- 
Catalin