Skip to contentSkip to conversation

Aliasing physical memory using virtual memory (from a d

DISCUSSION Petr Vandrovec 2 participants 4 messages Sep 18, 2003

lore
Share & tools
lore
#0 lore raw
Quoted 9 lines · Stevie-O

My thinking is this: I want to use __get_free_pages(1) 80 times to get the 160 pages, then passed those 80 pieces to the card (it's known the card can handle requests with that many pieces). Then I want to create a *virtually* contiguous 160-page mapping, so the postprocessing code in the driver can view the 80 2-page sub-buffers as one big consecutive 160-page buffer. Doing this would (a) make for more efficient use of memory, and (b) leave the larger piles of contiguous pages to the drivers of cards that actually require them.

If you'll use __get_free_pages(0) 160 times, you should be able to use vmap() in 2.[456].x.

I must say that I do not understand why it checks for
size > (max_mapnr << PAGE_SHIFT) in 2.4.x, or for count > num_physpages in 2.6.x (as there is nothing wrong with mapping same page several thousand times, or is it bad? with 32MB host you have plenty of unused VA space in the kernel...), but it should not hurt you as you need distinct physical pages.

On other side, maybe that using SG even for driver operations is not that complicated. Do not forget that on bigmem boxes you have only 128MB area for vmalloc/vmap/ioremap, so you can quickly find that there is not 640KB continuous area available.
Best regards,
Petr Vandrovec [email protected]

#1 lore raw
Quoted 11 lines · Petr Vandrovec
Quoted 7 lines · Stevie-O

My thinking is this: I want to use __get_free_pages(1) 80 times to get the
160 pages, then passed those 80 pieces to the card (it's known the card can
handle requests with that many pieces). Then I want to create a *virtually*
contiguous 160-page mapping, so the postprocessing code in the driver can
view the 80 2-page sub-buffers as one big consecutive 160-page buffer. Doing this would (a) make for more efficient use of memory, and (b) leave the larger piles of contiguous pages to the drivers of cards that actually
require them.

Quoted 2 lines

If you'll use __get_free_pages(0) 160 times, you should be able to use vmap() in 2.[456].x.

Actually, I specified __get_free_pages(1) 80 times because I don't know if the card's SG can actually support 160 separate buffers (I'm certain it can do at least 80 though).

I grepped my 2.4 kernel source for 'vmap' and the only results that seemed meaningful were vmap_pte_range or vmap_pmd_range in mips/mm/umap.c and mips64/mm/umap.c. Is this documented somewhere? I suffer from the 'i'm new at this, but this looks possible' syndrome. I don't actually know how anything is accomplished.

Btw, am I right about kmalloc(35000) effectively grabbing 64K?

Quoted 11 lines

I must say that I do not understand why it checks for size > (max_mapnr << PAGE_SHIFT) in 2.4.x, or for count > num_physpages in 2.6.x (as there is nothing wrong with mapping same page several
thousand times, or is it bad? with 32MB host you have plenty of
unused VA space in the kernel...), but it should not hurt you as you
need distinct physical pages.

On other side, maybe that using SG even for driver operations is not that complicated. Do not forget that on bigmem boxes you have only 128MB area for vmalloc/vmap/ioremap, so you can quickly find that there is not 640KB continuous area available.

Signature · 3 lines
- Stevie-O

Real Programmers use COPY CON PROGRAM.EXE
#2 lore raw
Quoted 4 lines · Stevie-O

I grepped my 2.4 kernel source for 'vmap' and the only results that seemed meaningful were vmap_pte_range or vmap_pmd_range in mips/mm/umap.c and mips64/mm/umap.c. Is this documented somewhere? I suffer from the 'i'm new at this, but this looks possible' syndrome. I don't actually know how anything is accomplished.

Btw, am I right about kmalloc(35000) effectively grabbing 64K?

I did a freetext search of the LXR for 'remap' and came up with this function:

820 /*

821  * maps a range of physical memory into the requested pages. the old
822  * mappings are removed. any references to nonexistent pages results
823  * in null mappings (currently treated as "copy-on-access")
824  */

825 static inline void remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
826 unsigned long phys_addr, pgprot_t prot)

This looks promising, although it's a bit ambiguous to me. Treating the physical pages as distinct from virtual pages:

[1] maps a range of physical memory onto the requested pages
I assume 'the requested pages' here refer to 'the specified virtual pages', since the statement makes no sense if it's referring to 'the specified physical pages'.

[2] the old mappings are removed.
Looking at the source, it seems that this means 'if these virtual pages were pointed to any physical pages before, they won't be anymore'

[3] any references to nonexistent pages results in null mappings ('copy-on-access')
I have no idea what this means, and I can't figure it out reading the code.
Obviously it refers to something not really being there (thus 'nonexistent'). So copy-on-access? How do you copy something that doesn't exist?

Signature · 3 lines
- Stevie-O

Real Programmers use COPY CON PROGRAM.EXE
#3 lore raw
Quoted 5 lines · Stevie-O
Quoted 3 lines · Stevie-O

I grepped my 2.4 kernel source for 'vmap' and the only results that seemed meaningful were vmap_pte_range or vmap_pmd_range in

2.4.22-ac1 has it. In mm/vmalloc.c. You cannot (well, I believe) use any functions which take page array if pages were not allocated one by one. Maybe you can try using page and page+1, but I'm under impression that it will not work as expected, and that you'll hit some BUG() somewhere.

Quoted 5 lines
Quoted 5 lines

mips/mm/umap.c and mips64/mm/umap.c. Is this documented somewhere? I suffer from the 'i'm new at this, but this looks possible' syndrome. I don't actually know how anything is accomplished.

Btw, am I right about kmalloc(35000) effectively grabbing 64K?

Yes.

Quoted 11 lines

I did a freetext search of the LXR for 'remap' and came up with this function:

820 /*

821  * maps a range of physical memory into the requested pages. the old
822  * mappings are removed. any references to nonexistent pages results
823  * in null mappings (currently treated as "copy-on-access")
824  */

825 static inline void remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
826 unsigned long phys_addr, pgprot_t prot)

Unavailable outside of mm. You must use remap_page_range. And this function can only remap memory which does not have its 'struct page' (i.e. MMIO on PCI busses) or pages marked as Reserved. So if you want to use it on regular memory, you must mark pages reserved... And then you have to do black magic to correctly remove 'PageReserved' bit at correct time - if process does fork, and you'll clear this bit too early, you'll get page_count < 0 and BUG(). If you'll do that too late, you'll leak memory. vmmon did this in the past, but it was impossible to get it right under all possible circumstances.

Other problem is that this functions is targeted for remapping userspace addresses, not kernel space, and I would not trust this function for using with from in kernel space. Definitely 2.6.x with 4G/4G patch will do bad things.
Best regards,
Petr Vandrovec [email protected]