Re: [PATCH] ppc: Add support for bigger page sizes than 4KB on PPC44x

On Fri Mar 16 16:14:48 EST 2007 Paul Mackerras wrote:

If I'm reading the patch right, with 64k pages you're using 8kB pgd
arrays and 256-byte pte arrays, so you're only using 1/256th of each
PTE page.  In fact each PTE page could map 512MB (assuming 8 byte
ptes), and the pgd array could be a tiny 8-entry thing, which we could
kmalloc.  I think that would use memory more efficiently.

 Your understanding of the patch is correct. When the page size is set to 64K 
then the 32-bit linear address is being splitted in the following way:

(1) 64K:
PGD: 11 msb -> there are 2K page global directory entries;
PTE: 5 bits -> 32 8-byte PTE entries in each page table => 32x8=256 of 64K;
Offset: 16 lsb -> page size is 64K.

 So, I do not use "page table" pages (with PTEs) fully indeed. If use the 
following decoding of the linear address then each page table with PTEs will 
be utilized in full:

(2) 64K:
PGD: 3 msb -> there are 8 page global directory entries,
PTE: 13 bits -> 8K 8-byte PTE entries in each page table => 8Kx8=64K of 64K,
Offset: 16 lsb -> page size is 64K.

 The corresponding point is correct regarding to my 16K implementation too:

(3) 16K:
PGD: 7 msb -> there are 128 page global directory entries,
PTE: 11 bits -> 2K 8-byte PTE entries in each page table => 2Kx8=16K of 16K,
Offset: 14 lsb -> page size is 16K.

 Additional note regarding to the 64K pages. When use decoding (2) there is a 
WARN_ON() triggered in the dma_alloc_init() function ( 
arch/powerpc/lib/dma-noncoherent.c):

WARN_ON(!pmd_none(*pmd)),

 that leads to the following message during kernel booting-up:
------------[ cut here ]------------
Badness at c02ebb30 [verbose debug info unavailable]
Call Trace:
[C054FE50] [C000858C] show_stack+0x48/0x190 (unreliable)
[C054FE80] [C00F189C] report_bug+0xa4/0xac
[C054FE90] [C0002A7C] program_check_exception+0x170/0x4a0
[C054FEB0] [C0001E74] ret_from_except_full+0x0/0x4c
[C054FF70] [C003C068] init_irq_proc+0x3c/0x5c
[C054FF80] [C02E0830] init+0x94/0x1c4
[C054FFF0] [C0003AC0] kernel_thread+0x44/0x60

 This warning means that the PMD, which is being checked here, was already 
allocated during executing of some prior code. When only 3 msb are being used 
as a PMD component of the linear address then each PMD entry covers the 512MB 
area (0x20000000). This is rather a big area so I think there is nothing 
strange or erroneous that this area was already accessed somewhere before the 
dma_alloc_init() call. So, to spare users' minds I suggest to replace the 
WARN_ON() call in the dma_alloc_init() with more harmless warning message. 

 The patch that fixes PTE tables memory usage for the 16K and 64K pages is 
attached. Also this patch replace WARN_ON() in the dma_alloc_init() with the 
printk(). 

 Regards, Yuri.