svn commit: r208589 - head/sys/mips/mips

[C. Jayachandran]

On Mon, Jun 7, 2010 at 10:57 PM, Alan Cox <alc at cs.rice.edu> wrote:
> C. Jayachandran wrote:
>>
>> On Fri, Jun 4, 2010 at 10:44 PM, Alan Cox <alc at cs.rice.edu> wrote:
>>
>>>
>>> C. Jayachandran wrote:
>>>
>>>>
>>>> On Thu, Jun 3, 2010 at 10:33 PM, Alan Cox <alc at cs.rice.edu> wrote:
>>>>
>>>>
>>>
>>> [snip]
>>>
>>>>>
>>>>> Add a static counter to pmap_ptpgzone_allocf().  When the counter
>>>>> reaches
>>>>> some not-too-small prime number, don't call vm_phys_alloc_contig(),
>>>>> instead
>>>>> set "m" to NULL and reset the counter to zero.  Setting "m" to NULL
>>>>> should
>>>>> then trigger the vm_contig_grow_cache() call.  Make sense?
>>>>>
>>>>>
>>>>
>>>> Is this to move the vm_contig_grow_cache() out of the
>>>> pmap_ptpgzone_allocf() and into the function calling uma_zalloc()?  I
>>>> am wondering why the prime number is needed.
>>>>
>>>> Another implementation I had thought about was to have a kernel
>>>> process maintain a pool of direct mapped pages for MIPS. This will be
>>>> woken up if the pool goes below a low water mark - any comments on
>>>> this?
>>>>
>>>>
>>>>
>>>
>>> Here is how the page table page allocation should be done.  It's not much
>>> harder than the vm_contig_grow_cache() change.
>>>
>>> 1. Look at vm/vm_phys.c, in particular, vm_phys_init().  Create a new
>>> freelist, like      VM_FREELIST_ISADMA, for the direct-mapped memory on
>>> MIPS.  Perhaps, call it VM_FREELIST_LOWMEM.  The controlling macros
>>> should
>>> be defined in mips/include/vmparam.h.
>>>
>>> 2. Trivially refactor vm_phys_alloc_pages().  Specifically, take the body
>>> of
>>> the outer for loop and make it a new function, vm_phys_alloc_freelist(),
>>>  that takes the variable "flind" as a parameter.
>>>
>>> 3. Eliminate the UMA zone for page table pages.  In place of
>>> vm_phys_alloc_contig() call your new function with VM_FREELIST_LOWMEM.
>>>  (The
>>> vm_contig_grow_cache() remains unchanged.)  Go back to calling
>>> vm_page_free() to release page table pages.
>>>
>>> For the record, the changes that you made to start using a zone for page
>>> table page allocation introduced at least one possible race condition
>>> between pmap_remove_pages() and pmap_remove_all().  Specifically, by
>>> dropping and reacquiring the page queues and pmap lock when you free a
>>> page
>>> table page, you allow a pmap_remove_all() call while pmap_remove_pages()
>>> is
>>> running to leave the variable "npv" in pmap_remove_pages() pointing at a
>>> freed pv entry.
>>>
>>
>> My first attempt is attached, it comes up multiuser but crashes if I
>> stress it a bit (panic: Bad link elm 0xc0078f00 prev->next != elm).
>> Will look at this tomorrow, and see if I can find the cause.
>>
>>
>
> Where exactly is this occurring?
>
>> In the meantime, it may be worth looking at the patch to see if this
>> is in line with the approach you had suggested. I have tried to  use
>> VM_FREELIST_HIGHMEM which is already there, instead of adding
>> VM_FREELIST_LOWMEM.
>>
>>
>
> Basically, yes.  At first, I was taken aback by defining VM_FREELIST_DEFAULT
> as other than zero, but it shouldn't break anything.

I can change this to add VM_FREELIST_LOWMEM. This works okay so far,
even though I haven't really checked the code to see if it works when
there is no memory in HIGHMEM.

> The one issue that I see with the patch is that you'll need to release and
> reacquire the pmap and page queues locks around the call to
> vm_contig_grow_cache().  However, if you release these locks, you need to
> restart the pmap_allocpte(), i.e., perform the "goto retry;", because while
> you slept another processor/thread might have already allocated and
> installed the page table page that you are trying to allocate into the page
> table.
>
> For what it's worth, this same race condition exists in the current code in
> subversion using an UMA zone, but not in the code that predates use of an
> UMA zone.

I have been testing with an updated version of the patch - my -j128
buildworld works on this. The sys/vm side changes are simple enough,
but the he page table page allocation part is still not complete. The
version I have now is (with gmail whitespace damage):

---
static vm_page_t
pmap_alloc_pte_page(pmap_t pmap, unsigned int index, int wait, vm_offset_t *vap)
{
       vm_paddr_t paddr;
       vm_page_t m;
       int tries, flags;

       tries = 0;
retry:
       mtx_lock(&vm_page_queue_free_mtx);
       m = vm_phys_alloc_freelist_pages(VM_FREELIST_DEFAULT,
           VM_FREEPOOL_DEFAULT, 0);
       if (m == NULL) {
               mtx_unlock(&vm_page_queue_free_mtx);
               if (tries < ((wait & M_NOWAIT) != 0 ? 1 : 3)) {
                       vm_contig_grow_cache(tries, 0, MIPS_KSEG0_LARGEST_PHYS);
                       tries++;
                       goto retry;
               } else {
                       printf("[%d] %s wait %x, fail!\n", tries, __func__,
                           wait);
                       return (NULL);
               }
       }

       m->pindex = index;
       m->valid = VM_PAGE_BITS_ALL;
       atomic_add_int(&cnt.v_wire_count, 1);
       m->wire_count = 1;
       if (m->flags & PG_ZERO)
               vm_page_zero_count--;
       else
               pmap_zero_page(m);
       flags = PG_ZERO | PG_UNMANAGED;
       m->flags = flags;
       m->oflags = 0;
       m->act_count = 0;
       if ((m->flags & PG_CACHED) != 0)
               printf("Not yet! handle cached  page %p flags 0x%x\n",
m, m->flags);
       else
               cnt.v_free_count--;

       m->act_count = 0;
       mtx_unlock(&vm_page_queue_free_mtx);
}


----

I tried to match the vm_page_alloc() code here. I'm not sure if I have
to handle the PG_CACHED case, and if it is okay to change the VM
counters here.

Also, I will change the code to  return NULL all the way and retry
allocation (similar to the old code which called VM_WAIT and then
returned NULL).

Probably I should get the whole code in line with the old code,
replacing the old vm_page_alloc() call with the function above without
retry, and the VM_WAIT with a new function which does the
vm_contig_grow_cache().

Thanks,
JC.