Buddy allocator - help! I thought I understood this

Buddy allocator - help! I thought I understood this

[Alex Bligh - linux-kernel]

I could swear I understood this bit of __free_pages_ok()
Monday night, but my mind appears to have gone blank.

As I recall, the memory bitmaps indicate by the
status a bit op returns whether or not a page
is on the free list for that particular order
area.

So how does this work (lines from vanilla 2.4.9
attached). I'm interested particularly in line
114:

  if (!__test_and_change_bit(index, area->map))

This examines the bitmap entry, in the
current order being examined bitmap.

As the page is being free'd, on the first pass,
won't the page ALWAYS register as unallocated?
This claims to test whether the buddy page
is still allocated. Wouldn't this test be:

  if (!__test_and_change_bit(index^1, area->map))

The annoying thing is I'm SURE I read through
this and understood it fully Monday night.

Or is the bitmap trying to tell us whether the
BUDDY is used? This doesn't seem to be what
expand et al are doing.

Please help - my head hurts.

(NB, if this is a code bug rather than local brain
 bug, this explains a lot)
--
Alex Bligh



    99          if (page_idx & ~mask)
   100                  BUG();
   101          index = page_idx >> (1 + order);
   102
   103          area = zone->free_area + order;
   104
   105          spin_lock_irqsave(&zone->lock, flags);
   106
   107          zone->free_pages -= mask;
   108
   109          while (mask + (1 << (MAX_ORDER-1))) {
   110                  struct page *buddy1, *buddy2;
   111
   112                  if (area >= zone->free_area + MAX_ORDER)
   113                          BUG();
   114                  if (!__test_and_change_bit(index, area->map))
   115                          /*
   116                           * the buddy page is still allocated.
   117                           */
   118                          break;
   119                  /*
   120                   * Move the buddy up one level.
   121                   */

Re: Buddy allocator - help! I thought I understood this

[Alex Bligh - linux-kernel]

thanks for the responses. FWIW:

> (NB, if this is a code bug rather than local brain
>  bug, this explains a lot)

It's was a transient local brain bug. The bit
is shared between two pages of the same order.

>    101          index = page_idx >> (1 + order);
                   I missed this today ^^^

And hence my previous patch did have a bug,
for arcane reasons; I'll resubmit it if
it works.

--
Alex Bligh

Re: Buddy allocator - help! I thought I understood this

[Linus Torvalds]

In article <525190103.999719280@[10.132.112.53]>,
Alex Bligh - linux-kernel  <linux-kernel@alex.org.uk> wrote:
>I could swear I understood this bit of __free_pages_ok()
>Monday night, but my mind appears to have gone blank.
>
>As I recall, the memory bitmaps indicate by the
>status a bit op returns whether or not a page
>is on the free list for that particular order
>area.

No. It actually indicates that the _buddy_ of the page is on the free
list for that particular order.

Basically each bit describes two pages of the order in question, which
is sufficient because while they have four states (both free, first
free, second free, neither free), we get the "other bit" of information
from the fact that we are just freeing one page - which obviously cannot
have been free before.

Thus the games with xor'ing the bit around, and looking at the old
state.

And no, most kernel programmers don't have to actually follow that piece
of code. Ugly but efficient.

		Linus