Re: [PATCH] mmc: block: delete packed command support

[Adrian Hunter <adrian.hunter@intel.com>]

On 21/11/16 16:02, Ulf Hansson wrote:
> On 21 November 2016 at 13:44, Adrian Hunter <adrian.hunter@intel.com> wrote:
>> On 21/11/16 13:11, Arnd Bergmann wrote:
>>> On Monday, November 21, 2016 11:08:57 AM CET Linus Walleij wrote:
>>>> I've had it with this code now.
>>>>
>>>> The packed command support is a complex hurdle in the MMC/SD block
>>>> layer, around 500+ lines of code which was introduced in 2013 in
>>>> commits
>>>>
>>>> ce39f9d17c14 ("mmc: support packed write command for eMMC4.5 devices")
>>>> abd9ac144947 ("mmc: add packed command feature of eMMC4.5")
>>>>
>>>> ...and since then it has been rotting. The original author of the
>>>> code has disappeared from the community and the mail address is
>>>> bouncing.
>>>>
>>>> For the code to be exercised the host must flag that it supports
>>>> packed commands, so in mmc_blk_prep_packed_list() which is called for
>>>> every single request, the following construction appears:
>>>>
>>>> u8 max_packed_rw = 0;
>>>>
>>>> if ((rq_data_dir(cur) == WRITE) &&
>>>>     mmc_host_packed_wr(card->host))
>>>>         max_packed_rw = card->ext_csd.max_packed_writes;
>>>>
>>>> if (max_packed_rw == 0)
>>>>     goto no_packed;
>>>>
>>>> This has the following logical deductions:
>>>>
>>>> - Only WRITE commands can really be packed, so the solution is
>>>>   only half-done: we support packed WRITE but not packed READ.
>>>>   The packed command support has not been finalized by supporting
>>>>   reads in three years!
>>>
>>> Packed reads don't make a lot of sense, there is very little
>>> for an MMC to optimize in reads that it can't already do without
>>> the packing. For writes, packing makes could be an important
>>> performance optimization, if the eMMC supports it.
>>>
>>> I've added Luca Porzio  and Alex Lemberg to Cc. I think they
>>> are subscribed to the list already, but it would be good to
>>> get some estimate from them too about how common the packed
>>> write support is on existing hardware from their respective
>>> employers before we kill it off.
>>>
>>> If none of Samsung/Micron/Sandisk are currently shipping
>>> devices that support eMMC-4.5 packed commands but don't
>>> support the eMMC-5.1 command queuing (which IIRC is a more
>>> general way to achieve the same), we probably don't need to
>>> worry about it too much.
>>>
>>>> - mmc_host_packed_wr() is just a static inline that checks
>>>>   host->caps2 & MMC_CAP2_PACKED_WR. The problem with this is
>>>>   that NO upstream host sets this capability flag! No driver
>>>>   in the kernel is using it, and we can't test it. Packed
>>>>   command may be supported in out-of-tree code, but I doubt
>>>>   it. I doubt that the code is even working anymore due to
>>>>   other refactorings in the MMC block layer, who would
>>>>   notice if patches affecting it broke packed commands?
>>>>   No one.
>>>
>>> This is a very good indication that it's not really used.
>>> It would be useful to also check out the Android AOSP
>>> kernel tree to see if it's in there, if anything important
>>> supports packed commands, it's probably in there.
>>>
>>>> - There is no Device Tree binding or code to mark a host as
>>>>   supporting packed read or write commands, just this flag
>>>>   in caps2, so for sure there are not any DT systems using
>>>>   it either.
>>>>
>>>> It has other problems as well: mmc_blk_prep_packed_list() is
>>>> speculatively picking requests out of the request queue with
>>>> blk_fetch_request() making the MMC/SD stack harder to convert
>>>> to the multiqueue block layer. By this we get rid of an
>>>> obstacle.
>>
>> SDHCIv4 has a feature (ADMA3) which is slightly similar to packed
>> commands but it does not require card support.
>>
>> Why is it a problem for blk-mq to allow some extra requests to
>> pick from for packing?
>>
> 
> In blk-mq, requests don't get picked from the queue, but instead those
> gets "pushed" to the block device driver.
> 
> First, to support packing, it seems like we would need to specify a
> queue-depth > 1, more or less lie to blk-mq layer about the device's
> capability. Okay, we can do that. But..
> 
> I also believe, the implementation would become really complex, as you
> would need to "hold" the first write request, while waiting for a
> second to arrive. Then for how long shall you hold it? And then what
> if you are unlucky so the next is read request, thus you can't pack
> them. The solution will be suboptimal, won't it?

It doesn't hold and wait now.  So why would it in the blk-mq case?