svn commit: r223307 - head/sys/vm

[Warner Losh]

On Jun 21, 2011, at 9:27 PM, Julian Elischer wrote:

> On 6/21/11 7:27 PM, Warner Losh wrote:
>> 
>> 
>> On Jun 21, 2011, at 5:27 PM, Alan Cox wrote:
>> 
>>> On 06/21/2011 16:09, Attilio Rao wrote:
>>>> 2011/6/21 Bruce Evans<brde at optusnet.com.au>:
>>>>> On Tue, 21 Jun 2011, Bjoern A. Zeeb wrote:
>>>>> 
>>>>>> On Jun 19, 2011, at 7:13 PM, Alan Cox wrote:
>>>>>> 
>>>>>> Hi Alan,
>>>>>> 
>>>>>>> Author: alc
>>>>>>> Date: Sun Jun 19 19:13:24 2011
>>>>>>> New Revision: 223307
>>>>>>> URL: http://svn.freebsd.org/changeset/base/223307
>>>>>>> 
>>>>>>> Log:
>>>>>>>  Precisely document the synchronization rules for the page's dirty field.
>>>>>>>  (Saying that the lock on the object that the page belongs to must be
>>>>>>> held
>>>>>>>  only represents one aspect of the rules.)
>>>>>>> 
>>>>>>>  Eliminate the use of the page queues lock for atomically performing
>>>>>>> read-
>>>>>>>  modify-write operations on the dirty field when the underlying
>>>>>>> architecture
>>>>>>>  supports atomic operations on char and short types.
>>>>>>> 
>>>>>>>  Document the fact that 32KB pages aren't really supported.
>>>>>> contrary to the tinderbox I'd like to point out that all mips kernels
>>>>>> built by universe are broken with a SVN HEAD from earlier today.  Could you
>>>>>> please check and see if you can fix it?  The errors I get are:
>>>>>> 
>>>>>> vm_page.o: In function `vm_page_clear_dirty':
>>>>>> /sys/vm/vm_page.c:(.text+0x18d0): undefined reference to `atomic_clear_8'
>>>>>> /sys/vm/vm_page.c:(.text+0x18d0): relocation truncated to fit: R_MIPS_26
>>>>>> against `atomic_clear_8'
>>>>>> vm_page.o: In function `vm_page_set_validclean':
>>>>>> /sys/vm/vm_page.c:(.text+0x38f0): undefined reference to `atomic_clear_8'
>>>>>> /sys/vm/vm_page.c:(.text+0x38f0): relocation truncated to fit: R_MIPS_26
>>>>>> against `atomic_clear_8'
>>>>> Atomic types shorter than int cannot be used in MI code, since they might
>>>>> not exist.  Apparently they don't exist on mips.  jake@ fixed all their
>>>>> old uses for sparc4 in ~Y2K.
>>>> I'm sure they do, they exist in support.S though and may not have the
>>>> _8 form (they may just have the _char version). I may look at the code
>>>> again to be sure.
>>>> 
>>> 
>>> It appears that while mips/include/atomic.h declares the existence of atomic_clear_8, mips/mips/support.S doesn't implement it.  In other words, only support for int's and short's is currently implemented, not char's:
>>> 
>>> # grep atomic_clear mips/mips/support.S
>>> * atomic_clear_32(u_int32_t *a, u_int32_t b)
>>> LEAF(atomic_clear_32)
>>> END(atomic_clear_32)
>>> * atomic_clear_16(u_int16_t *a, u_int16_t b)
>>> LEAF(atomic_clear_16)
>>> END(atomic_clear_16)
>> 
>> The current crop of atomic*16 and atomic*8 functions have the restriction that the address must be 32-bit aligned (and it forces this by aligning to 32-bits silently and then operates on the low 8 or 16 bits in that word!)
>> 
>> I'm guessing that this is likely just wrong.  Comments?
> 
> I'm guessing it depends on whether the hardware supports atomic sub-wordline accesses.
> (mind you it's getting really hard to work out what a wordline means these days)

Generally, they don't.  ll and lld are for 4 and 8 byte values.  You're supposed to load the word or doubleword with the instruction, mask off the byte you want to change, change it, and then write it back with the sc or scd insturction.  That will only succeed of none of the bytes in that word could have changed (yes, weasel words usually meaning anything in the cache line, but really can mean anything).

I know that some cores do support special instructions to do this, but they are with extensions to the ISA...

Warner