F18/rawhide: lockdep consumes a lot of percpu data

[Prarit Bhargava]

When booting F18 on a 32 cpu system I repeatedly see errors like:

PERCPU: allocation failed, size=320 align=32, alloc from reserved chunk failed
Pid: 620, comm: udevd Not tainted 3.7.0-0.rc5.git1.2.fc19.x86_64 #1
Call Trace:
 [<ffffffff81186c7d>] pcpu_alloc+0x9dd/0xa40
 [<ffffffff810e8b5a>] ? mod_verify_sig+0x39a/0x460
 [<ffffffff810e1f2e>] ? find_sec+0x5e/0x90
 [<ffffffff81186f83>] __alloc_reserved_percpu+0x13/0x20
 [<ffffffff810e5b5f>] load_module+0x40f/0x2290
 [<ffffffff812e68b2>] ? cred_has_capability+0x62/0x100
 [<ffffffff8109a333>] ? up_write+0x23/0x40
 [<ffffffff810e7a32>] sys_init_module+0x52/0x230
 [<ffffffff816fb799>] system_call_fastpath+0x16/0x1b
kvm: Could not allocate 320 bytes percpu data


This occurs because the lockdep code is allocating a giant chunk of memory in
the percpu data:

On 3.6.6-1:

14940 (D) __per_cpu_end

On 3.7.0-0:

1d5900 (D) __per_cpu_end

and in 3.7.0-0, a dump of the symbol map shows

efc0 (V) tstats_lookup_lock
f020 (V) cpu_lock_stats
1cef40 (V) stop_cpus_work
1cef80 (V) cpu_stopper

ie) the cpu_lock_stats takes up a huge chunk of the per_cpu memory.

I'm not sure what can be done about this; I suppose we could expand the reserved
area to allow for more module percpu allocation ...

P.

[Prarit Bhargava]

On 11/14/2012 03:51 PM, Prarit Bhargava wrote:
>  When booting F18 on a 32 cpu system I repeatedly see errors like:
>  
>  PERCPU: allocation failed, size=320 align=32, alloc from reserved chunk failed
>  Pid: 620, comm: udevd Not tainted 3.7.0-0.rc5.git1.2.fc19.x86_64 #1
>  Call Trace:
>   [<ffffffff81186c7d>] pcpu_alloc+0x9dd/0xa40
>   [<ffffffff810e8b5a>] ? mod_verify_sig+0x39a/0x460
>   [<ffffffff810e1f2e>] ? find_sec+0x5e/0x90
>   [<ffffffff81186f83>] __alloc_reserved_percpu+0x13/0x20
>   [<ffffffff810e5b5f>] load_module+0x40f/0x2290
>   [<ffffffff812e68b2>] ? cred_has_capability+0x62/0x100
>   [<ffffffff8109a333>] ? up_write+0x23/0x40
>   [<ffffffff810e7a32>] sys_init_module+0x52/0x230
>   [<ffffffff816fb799>] system_call_fastpath+0x16/0x1b
>  kvm: Could not allocate 320 bytes percpu data
>  
>  
>  This occurs because the lockdep code is allocating a giant chunk of memory in
>  the percpu data:
>  
>  On 3.6.6-1:
>  
>  14940 (D) __per_cpu_end
>  
>  On 3.7.0-0:
>  
>  1d5900 (D) __per_cpu_end
>  
>  and in 3.7.0-0, a dump of the symbol map shows
>  
>  efc0 (V) tstats_lookup_lock
>  f020 (V) cpu_lock_stats
>  1cef40 (V) stop_cpus_work
>  1cef80 (V) cpu_stopper
>  
>  ie) the cpu_lock_stats takes up a huge chunk of the per_cpu memory.
>  
>  I'm not sure what can be done about this; I suppose we could expand the reserved
>  area to allow for more module percpu allocation ...
>   
My explanation above isn't correct, however, the solution is likely the same.
I'll post on LKML ...

P.