LOR in VM (with backtrace)

[Alan L. Cox]

Kris Kennaway wrote:
> 
> CURRENT dated June 19;
> 
> lock order reversal
>  1st 0xc45788ac vm object (vm object) @ /a/asami/portbuild/i386/src-client/sys/vm/vm_object.c:1506
>  2nd 0xc082f110 system map (system map) @ /a/asami/portbuild/i386/src-client/sys/vm/vm_kern.c:328
> 
> Debugger(c03f450d,c082f110,c043190e,c043190e,c043178f) at Debugger+0x54
> witness_lock(c082f110,8,c043178f,148,d8ca79f8) at witness_lock+0x6ac
> _mtx_lock_flags(c082f110,0,c043178f,148,c4496850) at _mtx_lock_flags+0xb1
> _vm_map_lock(c082f0b0,c043178f,148,c025dc34,246) at _vm_map_lock+0x36
> kmem_malloc(c082f0b0,1000,101,d8ca7a8c,c0390425) at kmem_malloc+0x65
> page_alloc(c083a240,1000,d8ca7a7f,101,c0457e2c) at page_alloc+0x27
> slab_zalloc(c083a240,101,c04332aa,66e,c414e5c4) at slab_zalloc+0x155
> uma_zone_slab(c083a240,101,c04332aa,66e,0) at uma_zone_slab+0xd8
> uma_zalloc_internal(c083a240,0,101,6ee,0) at uma_zalloc_internal+0x55
> uma_zfree_arg(c414e5a0,d7a86000,0,1,0) at uma_zfree_arg+0x2cc
> swp_pager_meta_free(c45788ac,1d,0,1,0) at swp_pager_meta_free+0x1b2
> swap_pager_freespace(c45788ac,1d,0,1,0) at swap_pager_freespace+0x57
> vm_object_backing_scan(c4bec5c8,4,c0432299,5f7,c0432299) at vm_object_backing_scan+0x28a
> vm_object_collapse(c4bec5c8,0,c0432299,1ef,c45a0ec4) at vm_object_collapse+0x11a
> vm_object_deallocate(c4fab940,c5093ce4,c4fab940,c5093ce4,d8ca7c60) at vm_object_deallocate+0x28e
> vm_map_entry_delete(c1506100,c5093ce4,c043197c,86d,c041ada4) at vm_map_entry_delete+0x3b
> vm_map_delete(c1506100,0,bfc00000,c1506100,c4123f00) at vm_map_delete+0x413
> vm_map_remove(c1506100,0,bfc00000,11d,65) at vm_map_remove+0x55
> exit1(c4496850,0,c041a26b,65,d8ca7d40) at exit1+0x63d
> sys_exit(c4496850,d8ca7d10,c0437c26,3fd,1) at sys_exit+0x41
> syscall(2f,2f,2f,0,ffffffff) at syscall+0x26e
> Xint0x80_syscall() at Xint0x80_syscall+0x1d
> --- syscall (1, FreeBSD ELF32, sys_exit), eip = 0x2829e9ff, esp = 0xbfbfde3c, ebp = 0xbfbfde58 ---
> 

Thanks for letting me know.  This is another false positive: Witness
can't distinguish the lock on the object being destroyed from the lock
on the object used by UMA because their labels are the same.  They will
never, however, be the same object.  So, deadlock isn't a risk.

I will probably disable witness checking on the "system map" mutexes (or
at least the kmem_map's mutex) shortly so that folks don't become
conditioned to ignoring these.

Regards,
Alan