Re: [PATCH] pci: print out DMA mask info

[Yinghai Lu <yinghai@kernel.org>]

FUJITA Tomonori wrote:
> On Thu, 9 Oct 2008 21:56:33 -0700
> "Yinghai Lu" <yinghai@kernel.org> wrote:
> 
>> On Thu, Oct 9, 2008 at 7:40 PM, FUJITA Tomonori
>> <fujita.tomonori@lab.ntt.co.jp> wrote:
>>> On Thu, 09 Oct 2008 16:05:57 -0700
>>> Yinghai Lu <yinghai@kernel.org> wrote:
>>>
>>>> Grant Grundler wrote:
>>>>> On Thu, Oct 09, 2008 at 02:51:32PM -0700, Yinghai Lu wrote:
>>>>>> On Thu, Oct 9, 2008 at 2:35 PM, Matthew Wilcox <matthew@wil.cx> wrote:
>>>>>>> Why's that interesting to the sysadmin of the machine?  To the driver
>>>>>>> writer, certainly.  But what's the use of it to the people using the
>>>>>>> machine?
>>>>> ...
>>>>>> make linux kernel act like black box as other os?
>>>>> I don't understand your reply.
>>>>> If someone thinks linux is a black box, printing this message won't help them.
>>>>>
>>>> could find out easily why some driver doesn't set dma mask correctly.
>>>> like why
>>>>       qlogic qla2xxx only set consistent to 64bit,
>>>>       emulex lpfc not set consistent to 64bit
>>> IIRC, except for one SGI architecture, coherent_dma_mask is
>>> meaningless, dma_mask is always equal to coherent_dma_mask. Lots of
>>> IOMMU implementations ignore coherent_dma_mask and use dma_mask for
>>> alloc_coherent(). Some drivers doesn't set up coherent_dma_mask.
>> ehci_hcd 0000:00:02.1: using 31bit consistent DMA mask
>> ==> ck804 ehci, is using 31bit for consistent dma mask, at still use
>> 32 bit for dma mask.
> 
> ehci_hcd needs to set 31bit to dma_mask, I guess.

in ehci_pci_setup() 

        switch (pdev->vendor) {
        case PCI_VENDOR_ID_NVIDIA:
                /* NVidia reports that certain chips don't handle
                 * QH, ITD, or SITD addresses above 2GB.  (But TD,
                 * data buffer, and periodic schedule are normal.)
                 */
                switch (pdev->device) {
                case 0x003c:    /* MCP04 */
                case 0x005b:    /* CK804 */
                case 0x00d8:    /* CK8 */
                case 0x00e8:    /* CK8S */
                        if (pci_set_consistent_dma_mask(pdev,
                                                DMA_31BIT_MASK) < 0)
                                ehci_warn(ehci, "can't enable NVidia "
                                        "workaround for >2GB RAM\n");
                        break;
                }
                break;
        }

so that is strange silicon bug for old ck804 and before. dma_mask could be 32bit. but consistent_dma_mask is 31bit
mcp55 is ok.


> 
> 
>> qlogic qla2xxx  and emulex lpfc dma mask and consistent_dma_mask is different...
>> could have some story for them
> 
> Check out qla2xxx again. I think that it uses dma_set_mask() to set
> dma_mask. qla2xxx uses the same value for both dma_mask and
> consistent_dma_mask.
> 
> lpfc had better set 64bit to consistent_dma_mask but as I said in the
> previous mail, not setting consistent_dma_mask doesn't cause any
> problem. It means that some IOMMUs (uses consistent_dma_mask properly)
> allocates an address < 4GB in alloc_coherent() and some IOMMUs alloc
> address > 4GB. lpfc can handle both anyway.
> 
> 
>> at least gart iommu is honoring the consistent dma mask.
>> by calling dma_alloc_coherent_mask(dev, flag)
> 
> Well, that's because I wrote gart's alloc_coherent and introduced
> dma_alloc_coherent_mask. ;)

oh. it is in tip

> 
> 
>> if device could use 64 bit coherent dma mask, that is driver problem...
> 
> Can you be more specific? As I wrote above, if 64bit-dma-capable
> devices don't set consistent_dma_mask, we don't have any problem.

then can we remove consistent_dma_mask? just use dma_mask instead for all.

> 
> Yes, drivers that have dma_mask < 32bit need to set up
> consistent_dma_mask. But few devices have dma_mask < 32bit and we can
> fix them without the information at boot time, I think.

YH