On 3 July 2018 at 20:45, Belisko Marek <marek.belisko@xxxxxxxxx> wrote:
> Hi Ulf,
>
> CC'ed pyra kernel guys
>
> On Tue, Jul 3, 2018 at 11:23 AM Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
>>
>> On 3 July 2018 at 08:47, Belisko Marek <marek@xxxxxxxxxxxxx> wrote:
>> > Hello,
>> >
>> > on pyra device [0] we're using txs02612 sdio expander to have
>> > possibility to connect SD card and also eMMC. We have some version of
>> > driver (gpio for switching bus is controlled via sysfs) which probably
>> > won't accepted. We want to do it correct way so we was thinking about
>> > same approach as i2c-mux is using (i2c device connected on mux buses
>> > are transparent to user so switching is done completely in driver). I
>> > would like to maybe get some initial thoughts if this is good idea to
>> > go this way and don't be a problem for future upstreaming. Thanks for
>> > any comments.
>>
>> Sorry, but I doubt we ever get to add proper upstream support for mmc
>> hosts managing multiple slots. Simply because all the efforts it takes
>> isn't worth it.
>>
>> Today there are some host trying to support this, but those are hacks,
>> which is working very poorly. I don't want to maintain more of those,
>> then it's better those are managed by vendors out of tree.
>
> Thanks for quick reply.
>
> (just copy part of internal discussion):
> My approach would be that the driver registers two new host
> controllers, e.g. like mmc1 and mmc2 on the omap5.
>
> I.e. calls in its probe function
>
> mmc[i] = mmc_alloc_host(sizeof(struct omap_hsmmc_host), &pdev->dev);
> mmc[i]->ops = txs02612_ops;
> // other setup
> mmc_add_host(mmc[i]);
>
> twice for each of the two txs channels. This should create two new
> user-space mmc ports that receive requests and other callbacks.
>
> It then should takes control of the real host controller to sit in
> the middle.
>
> So basically we should do something like:
>
> host_mmc = sarch_by_of_handle(...)
>
> And then make the txs02612_ops like:
>
> txs02612_request(mmc, ...) {
> find out which port
> if(different from current)
> {
> try to get mutex
> throw the switch
> }
> host_mmc->request(host_mmc, ...)
> if(was different)
> release the mutex
> }
>
> in this case we don't need to touch mmc core and also bus locking will
> be done in driver.
> Do you think we can get some success with upstreaming by this approach?
No.
> Thanks.
>>
>> There have been several discussions on this topic, this is just one
>> that I found while searching.
>> https://lkml.org/lkml/2017/10/6/710
Did you read this?
>>
>> If you need any further explainations, please just ask.
>>
Let me elaborate a bit on the problems.
1. Serialization of requests (several request sent after each other,
with delays, changing clocks etc), is managed by the core, because it
implements the SD/SDIO/(e)MMC specs. This does not belong in a host
driver.
-> A kind of mmc bus slot lock is needed in the core. Not a big deal,
we can probably hide it below mmc_claim|release_host().
2. What if there is an SD card that can cope with high-speed mode in
one slot with a clock freq at 50 MHz. Then in the other slot, there is
an eMMC using HS200 mode, running at 200 MHz. Can the HW guarantee all
kind of different combinations, with clock glitches etc?
3. The biggest problem: You can *not* provide any service guarantees
in regards to latency or bandwidth for any of the storage media in the
slots, because of the bus slot lock. That because the upper layers
runs I/O scheduling separately for each storage device, thus request
for one slot can starve requests on another.
Kind regards
Uffe
--
To unsubscribe from this list: send the line "unsubscribe linux-mmc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
