From: Florian Westphal <f...@strlen.de>
Date: Fri, 12 Aug 2016 04:13:43 +0200
> Sander reports following splat after netfilter nat bysrc table got
> converted to rhashtable:
>
> swapper/0: page allocation failure: order:3,
> mode:0x2084020(GFP_ATOMIC|__GFP_COMP)
> CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.8.0-rc1 [..]
> [<ffffffff811633ed>] warn_alloc_failed+0xdd/0x140
> [<ffffffff811638b1>] __alloc_pages_nodemask+0x3e1/0xcf0
> [<ffffffff811a72ed>] alloc_pages_current+0x8d/0x110
> [<ffffffff8117cb7f>] kmalloc_order+0x1f/0x70
> [<ffffffff811aec19>] __kmalloc+0x129/0x140
> [<ffffffff8146d561>] bucket_table_alloc+0xc1/0x1d0
> [<ffffffff8146da1d>] rhashtable_insert_rehash+0x5d/0xe0
> [<ffffffff819fcfff>] nf_nat_setup_info+0x2ef/0x400
>
> The failure happens when allocating the spinlock array.
> Even with GFP_KERNEL its unlikely for such a large allocation
> to succeed.
>
> Thomas Graf pointed me at inet_ehash_locks_alloc(), so in addition
> to adding NOWARN for atomic allocations this also makes the bucket-array
> sizing more conservative.
>
> In commit 095dc8e0c3686 ("tcp: fix/cleanup inet_ehash_locks_alloc()"),
> Eric Dumazet says: "Budget 2 cache lines per cpu worth of 'spinlocks'".
> IOW, consider size needed by a single spinlock when determining
> number of locks per cpu.
>
> Currently, rhashtable just allocates 128 locks per cpu which gives
> factor of 4 more than what inet hashtable uses with same number of
> cpus.
>
> For LOCKDEP, we now allocate a lot less locks than before (1 per cpu on
> my test box) so we no longer need to pretend we only have two cpus.
>
> Some sizes (64 byte L1 cache, 4 byte per spinlock, numbers in bytes):
>
> cpus: 1 2 4 8 16 32 64
> old: 1k 1k 4k 8k 16k 16k 16k
> new: 128 256 512 1k 2k 4k 8k
>
> With 72-byte spinlock (LOCKDEP):
> cpus : 1 2 4 8 16 32 64
> old: 9k 18k 18k 18k 18k 18k 18k
> new: 72 144 288 575 ~1k ~2.3k ~4k
>
> Reported-by: Sander Eikelenboom <li...@eikelenboom.it>
> Suggested-by: Thomas Graf <tg...@suug.ch>
> Signed-off-by: Florian Westphal <f...@strlen.de>
Applied, thanks Florian.
