malloc.9 locking section
John Baldwin
jhb at FreeBSD.org
Wed Apr 9 09:00:25 PDT 2003
On 09-Apr-2003 Alan Cox wrote:
> On Tue, Apr 08, 2003 at 04:47:14PM -0400, John Baldwin wrote:
>>
>> On 08-Apr-2003 Alan Cox wrote:
>> > On Tue, Apr 08, 2003 at 03:31:40PM -0400, John Baldwin wrote:
>> >>
>> >> On 17-Mar-2003 Harti Brandt wrote:
>> >> > Index: malloc.9
>> >> > ===================================================================
>> >> > RCS file: /home/ncvs/src/share/man/man9/malloc.9,v
>> >> > retrieving revision 1.30
>> >> > diff -u -r1.30 malloc.9
>> >> > --- malloc.9 24 Feb 2003 05:53:27 -0000 1.30
>> >> > +++ malloc.9 17 Mar 2003 15:06:14 -0000
>> >> >
>> >> > [snip]
>> >>
>> >> Looks good to me. While you are at it, please kill the following
>> >> from the manpage (if you aren't already doing so):
>> >>
>> >> Any calls to malloc() or free() when holding a vnode(9) interlock, will
>> >> cause a LOR (Lock Order Reversal) due to the interwining of VM Objects
>> >> and Vnodes.
>> >>
>> >
>> > Why? The above statement is true.
>>
>> It's highly specific. Harti is adding wording to say "don't hold any
>> locks when calling malloc() with M_WAITOK," not just vnode interlocks.
>> If vnode interlocks are even a problem with M_NOWAIT, then perhaps you
>> could add wording for that case to Harti's statement ("even with M_NOWAIT
>> one cannot hold vnode interlocks...").
>
> Holding a vnode interlock is problematic regardless of whether M_WAITOK
> or M_NOWAIT is specified. It's a rather non-obvious special case. Even
> free() is problematic. In December or January, I recall there being
> several reported lock order reversals due to this. This inspired someone
> to add the above comment to the man page.
>
>> ... My main concern is that I don't want
>> a situation where malloc(9) grows a huge laundry list of all the locks
>> in the kernel saying that can't be held when it is called. Such a list
>> would be hard to maintain and would easily rot, be incomplete, etc.
>>
>
> I think this is a one-of-a-kind special case. The vnode interlock is
> the only lock in this part of the memory management system that gets
> shared with another part of the kernel.
Ok.
--
John Baldwin <jhb at FreeBSD.org> <>< http://www.FreeBSD.org/~jhb/
"Power Users Use the Power to Serve!" - http://www.FreeBSD.org/
More information about the freebsd-smp
mailing list