On Thu, Feb 15, 2007 at 08:29:28PM +0100, Blaisorblade wrote:
> The following could be a suggestion, if max_low_pfn is not used between the
> old and the new moment of assignment (and it seems it is not). This is just
> an idea however:
> - max_low_pfn = ...
> /* this will put all low memory onto the freelists */
> totalram_pages = free_all_bootmem();
> + max_low_pfn = totalram_pages;
> #ifdef CONFIG_HIGHMEM
> totalhigh_pages = highmem >> PAGE_SHIFT;
> totalram_pages += totalhigh_pages;
> num_physpages = totalram_pages;
> max_pfn = totalram_pages;
> Please note that I did not spend a lot of time on this, so everything
> could be wrong. However, testing cannot help with uml_reserved
> handling, and this is a dark corner. So things should be better
> understood before merging the patch.
I agree - I have only a vague idea about what uml_reserved means.
> The code is too convoluted for a brief look - drawing a picture which
> explains all those variables would help. Both for UML and for every
blk_queue_bounce_limit() calls init_emergency_isa_pool() to get dma-zone
pages to use as bounce buffers when its caller passes a dma_addr limit
that is less than max_low_pfn. The BLK_BOUNCE_ANY macro is supposed to
mean "never bounce", and it's defined as:
#define BLK_BOUNCE_ANY ((u64)blk_max_pfn << PAGE_SHIFT)
So it presumes that max_pfn >= max_low_pfn. uml's mem_init() is
violating this assumption - when uml_reserved is subtracted from
max_pfn, we end up with max_pfn < max_low_pfn, so BLK_BOUNCE_ANY has the
opposite of the intended effect. blk_queue_bounce_limit therefore tries
to create a mempool with zone-dma pages on a no-dma-zone arch and the
kernel goes BUG().
So I think your idea is correct. It passes my testing - I can still use
lvm within uml. I have not tested CONFIG_HIGHMEM, but here's an
implementation against 2.6.20.
Jeff, please drop my other patch and use this one.
Signed-off-by: Jason Lunz <lunz@...>
arch/um/kernel/mem.c | 3 +--
1 file changed, 1 insertion(+), 2 deletions(-)
@@ -63,8 +63,6 @@
- max_low_pfn = (high_physmem - uml_physmem) >> PAGE_SHIFT;
/* clear the zero-page */
memset((void *) empty_zero_page, 0, PAGE_SIZE);
@@ -79,6 +77,7 @@
/* this will put all low memory onto the freelists */
totalram_pages = free_all_bootmem();
+ max_low_pfn = totalram_pages;
totalhigh_pages = highmem >> PAGE_SHIFT;
totalram_pages += totalhigh_pages;