Re: bus_space_physload(9)

[Masao Uebayashi <uebayasi%tombi.co.jp@localhost>]

On Wed, Mar 24, 2010 at 01:07:15PM -0500, David Young wrote:
> Can it be implemented like this?
> 
> void *
> bus_space_physload(bus_space_tag_t bst, bus_addr_t addr, bus_size_t size,
>     vm_prot_t prot, int flags)
> {
>       bus_space_handle_t bsh;
>       void *p;
>       int rc;
> 
>       rc = bus_space_map(bst, addr, size, flags | BUS_SPACE_MAP_LINEAR, &bsh);
>       if (rc != 0)
>               return NULL;
> 
>       p = bus_space_vaddr(bst, bsh);
> 
>       if (p == NULL) {
>               bus_space_unmap(bst, bsh, size);
>               return NULL;
>       }
> 
>       rc = vm_map_protect(kernel_map, (vaddr_t)p, (vaddr_t)p + size, prot,
>           true);
> 
>       if (rc != 0) {
>               bus_space_unmap(bst, bsh, size);
>               return NULL;
>       }
> 
>       return p;
> }

No, like this:

        void *
        bus_space_physload(bus_space_tag_t t, bus_addr_t addr, bus_size_t size,
            int prot, int flags)
        {
                const bus_addr_t busp = addr;
                const bus_addr_t busq = addr + size;
                const paddr_t physp = bus_space_mmap(t, busp, 0, prot, flags);
                const paddr_t physq = bus_space_mmap(t, busq, 0, prot, flags);

                /* return struct vm_physseg * as a cookie */
                return uvm_page_physload_device(
                        physp,          /* physical start */
                        physq);         /* physical end */
        }

bus_space_physload(9) registers a physical address region into VM, with
protection and some properties (vm_memattr_t in FreeBSD, or enum pmem_props
in your pmem(9) proposal).  Probably it'll have struct vm_page_md for
VM_PROT_EXEC regions for XIP.

I've decided to *not* allocate unmanaged regions here.  Unmanaged regions
are tracked by bus_space(9) tags / handles.  They're always mapped read/write,
and uncached.  No need to keep their states.

Eventually we'll return bus_space_physload(9) cookie back to cdev_mmap(9),
which in turn passes the cookie to pmap_enter(9).  This means that we can
pass device's physical address and its properties.

bus_dmamem_mmap(9) and bus_space_mmap(9) exposing paddr_t is wrong.
cdev_mmap(9) should return either "struct vm_page *" (bus_dmamem_mmap(9) ==
DMA memory) or "struct vm_physseg *" + "off_t" (bus_space_physload(9) ==
device region).

Masao

-- 
Masao Uebayashi / Tombi Inc. / Tel: +81-90-9141-4635