virt_to_page(pci_alloc_consistent())

Guys,

I just noticed that sound drivers use the address from pci_alloc_consistent() as the input to virt_to_page() all over the place. I looked into the Documentation/DMA-mapping.txt, and it says:

This routine will allocate RAM for that region, so it acts similarly to __get_free_pages (but takes size instead of a page order).

I know for fact I got it wrong in sparc in whole 2.4, and it seems RMK got it wrong in arm. I suggest other architecture maintainers to look at it ASAP. May even be oopsabe, by indexing outside of mem_map[] with a suitable sound driver.

-- Pete

From: Pete Zaitcev <[email protected]>
Date: Thu, 3 Oct 2002 02:38:14 -0400

I just noticed that sound drivers use the address from pci_alloc_consistent() as the input to virt_to_page() all over the place. I looked into the Documentation/DMA-mapping.txt, and it says:

This routine will allocate RAM for that region, so it acts similarly to __get_free_pages (but takes size instead of a page order).

I know for fact I got it wrong in sparc in whole 2.4, and it seems RMK got it wrong in arm. I suggest other architecture maintainers to look at it ASAP. May even be oopsabe, by indexing outside of mem_map[] with a suitable sound driver.

I think we MUST allow architectures to do what SPARC and ARM do, there is simply no other way to change the page protections easily for these kinds of systems.

So please instead change the DMA-mapping.txt documentation and propose pci_consistent_to_page(ptr) API additions. Next, inform the subsytems trying to use virt_to_page() on this memory that they need to use the new interface.

I can fix sparc, dunno about others. Do you have a problem on sparc64? Are you using a huge page to map the kernel?

It's hell on earth trying to find them without __pa barfing on addresses outside of linear area. Don't you think it may be time to add a BUG_ON there, at least for 2.5?

I also toyed with idea of __pa which walks page tables. The problem now is that every Joe Driver Writer on the street started to use pte_offset(pmd_offset(pgd_offset())), which they duly copy from bttv. Of course, highmem or other changes break that, because it was not an API meant to be exposed. We kinda let Kraxel to do it because there was no other way for V4L.

The idea of some virtual addresses more equial than others is extremely hard to impart into heads of driver writers, in my expirience. At least we have to invent a new term for vmalloc area, such as "logical" address, and make Corbet to fix THE BOOK.

-- Pete

From: Pete Zaitcev <[email protected]>
Date: Thu, 3 Oct 2002 12:25:10 -0400

I can fix sparc, dunno about others. Do you have a problem on sparc64? Are you using a huge page to map the kernel?

ARM I believe does not have luxury you have.

See, some processors are like MIPS and something sort of KSEG0 is used for PAGE_OFFSET mapping of physical pages. So pages with page protections changes must live in some other area of the kernel address space just like vmalloc() to get the different protection bits.

I really would like sun4m to use big pages for the PAGE_OFFSET mappings when possible Peter. So I even don't think fixing Sparc the way I know you are likely to is wise.