PATCH: fix bogus error message "bus_dmamem_alloc failed to align memory properly"

[Neel Natu]

Hi John,

On Tue, Sep 28, 2010 at 6:36 AM, John Baldwin <jhb at freebsd.org> wrote:
> On Monday, September 27, 2010 5:13:03 pm Neel Natu wrote:
>> Hi John,
>>
>> Thanks for reviewing this.
>>
>> On Mon, Sep 27, 2010 at 8:04 AM, John Baldwin <jhb at freebsd.org> wrote:
>> > On Friday, September 24, 2010 9:00:44 pm Neel Natu wrote:
>> >> Hi,
>> >>
>> >> This patch fixes the bogus error message from bus_dmamem_alloc() about
>> >> the buffer not being aligned properly.
>> >>
>> >> The problem is that the check is against a virtual address as opposed
>> >> to the physical address. contigmalloc() makes guarantees about
>> >> the alignment of physical addresses but not the virtual address
>> >> mapping it.
>> >>
>> >> Any objections if I commit this patch?
>> >
>> > Hmmm, I guess you are doing super-page alignment rather than sub-page
>> > alignment?  In general I thought the busdma code only handled sub-page
>> > alignment and doesn't fully handle requests for super-page alignment.
>> >
>>
>> Yes, this is for allocations with sizes greater than PAGE_SIZE and
>> alignment requirements also greater than a PAGE_SIZE.
>>
>> > For example, since it insists on walking individual pages at a time, if you
>> > had an alignment setting of 4 pages and passed in a single, aligned 4-page
>> > buffer, bus_dma would actually bounce the last 3 pages so that each individual
>> > page is 4-page aligned.  At least, I think that is what would happen.
>> >
>>
>> I think you are referring to bus_dmamap_load() operation that would
>> follow the bus_dmamem_alloc(), right? The memory allocated by
>> bus_dmamem_alloc() does not need to be bounced. In fact, the dmamap
>> pointer returned by bus_dmamem_alloc() is NULL.
>>
>> At least for the amd64 implementation there is code in
>> _bus_dmamap_load_buffer() which will coalesce individual dma segments
>> if they satisfy 'boundary' and 'segsize' constraints.
>
> So the problem is earlier in the routine where it does this:
>
>                /*
>                 * Get the physical address for this segment.
>                 */
>                if (pmap)
>                        curaddr = pmap_extract(pmap, vaddr);
>                else
>                        curaddr = pmap_kextract(vaddr);
>
>                /*
>                 * Compute the segment size, and adjust counts.
>                 */
>                max_sgsize = MIN(buflen, dmat->maxsegsz);
>                sgsize = PAGE_SIZE - ((vm_offset_t)curaddr & PAGE_MASK);
>                if (map->pagesneeded != 0 && run_filter(dmat, curaddr)) {
>                        sgsize = roundup2(sgsize, dmat->alignment);
>                        sgsize = MIN(sgsize, max_sgsize);
>                        curaddr = add_bounce_page(dmat, map, vaddr, sgsize);
>                } else {
>                        sgsize = MIN(sgsize, max_sgsize);
>                }
>
> If you have a map that does need bouncing, then it will split up the pages.
> It happens to work for bus_dmamem_alloc() because that returns a NULL map
> which doesn't bounce.  But if you had a PCI device which supported only
> 32-bit addresses on a 64-bit machine with an aligned, 4 page buffer above
> 4GB and did a bus_dma_map_load() on that buffer, it would get split up into
> 4 separate 4 page-aligned pages.
>

You are right.

I assume that you are ok with the patch and the discussion above was
an FYI, right?

best
Neel

> --
> John Baldwin
>