Benjamin Herrenschmidt wrote:
Hi !
I'd like to kick a discussion about some issues I've been having along
with some proposed solutions related to mapping of bits of hardware in
smarter ways than simply doing a io_remap_pfn_range() and the problems,
generally caused by get_user_pages().
...
The Tungstengrpahics folks (Thomas is on CC) have been working on some
better memory management to properly handle those things in the DRM. One
of the things we want to do here is similar to what the SPUs do with
local store: have a single VMA associated with an object, and have the
PTEs transparently changed to map either video memory, system memory,
AGP memory, etc... (the different in cache attributes can be ignored at
this stage, we can discuss it separately if interested).
I've been suggesting a similar approach as we use for SPUs. That is what
would make the most sense from a user standpoint: user code access their
"objects" via a single virtual pointer and the DRM takes care of
migrating it when necessary (for example migrating it to video RAM when
it needs to be used by the engine and "swap" it back to main memory when
not).
...
The base idea is that we would have the no_page() function of SPU's or
the DRM either return a struct page when the object is backed to main
memory, or install the PTE directly (using the helper to hide some of
the low level TLB flushing logic etc...) and then return NOPAGE_REFAULT
when hitting the hardware. The helper basically is a one-page version of
io_remap_pfn_range() with the added "feature" of not doing anything if
the PTE has been set by somebody else (handle the race case) instead of
BUG'ing as the current io_remap_pfn_range() does.
I'm very much for this approach, possibly with the extension that we
could have a multiple-page version as well, as populating the whole vma
sometimes may be cheaper than populating each pte with a fault. That
would basically be an io_remap_pfn_range() which is safe when the
mmap_sem is taken in read mode (from do_no_page).
One problem that occurs is that the rule for ptes with non-backing
struct pages
Which I think was introduced in 2.6.16:
pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
cannot be honored, at least not with the DRM memory manager, since the
graphics object will be associated with a vma and not the underlying
physical address. User space will have vma->vm_pgoff as a handle to the
object, which may move around in graphics memory.
/Thomas
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[Index of Archives]
[Kernel Newbies]
[Netfilter]
[Bugtraq]
[Photo]
[Stuff]
[Gimp]
[Yosemite News]
[MIPS Linux]
[ARM Linux]
[Linux Security]
[Linux RAID]
[Video 4 Linux]
[Linux for the blind]
[Linux Resources]
