Message-ID: <CAPDyKFrtVV35mg4dq5UZRnAzNRQRAbmbRXQfCB_gdpdyPETK3Q@mail.gmail.com>
Date:   Mon, 28 Mar 2022 13:38:25 +0200
From:   Ulf Hansson <ulf.hansson@...aro.org>
To:     Michael Wu <michael@...winnertech.com>
Cc:     Adrian Hunter <adrian.hunter@...el.com>,
        Christian Löhle <CLoehle@...erstone.com>,
        Avri Altman <avri.altman@....com>,
        "beanhuo@...ron.com" <beanhuo@...ron.com>,
        "porzio@...il.com" <porzio@...il.com>,
        "linux-mmc@...r.kernel.org" <linux-mmc@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        allwinner-opensource-support 
        <allwinner-opensource-support@...winnertech.com>
Subject: Re: [PATCH] mmc: block: enable cache-flushing when mmc cache is on

On Mon, 28 Mar 2022 at 12:11, Michael Wu <michael@...winnertech.com> wrote:
>
> On 25/03/2022 18:13, Ulf Hansson wrote:
> > On Fri, 25 Mar 2022 at 06:46, Michael Wu <michael@...winnertech.com> wrote:
> >>
> >> On 24/03/2022 19:27, Ulf Hansson wrote:
> >>> On Thu, 17 Mar 2022 at 10:14, Ulf Hansson <ulf.hansson@...aro.org> wrote:
> >>>>
> >>>> On Wed, 16 Mar 2022 at 17:08, Adrian Hunter <adrian.hunter@...el.com> wrote:
> >>>>>
> >>>>> On 16.3.2022 16.46, Christian Löhle wrote:
> >>>>>>> So we are not going to let the block layer know about SD cache?
> >>>>>>> Or is it a separate change?
> >>>>>>
> >>>>>> I have some code for this laying around, but as it requires reading, parsing and writing Function Registers,
> >>>>>> in particular PEH, it's a lot of boilerplate code to get the functionality, but I'll clean it up and send a patch in the coming weeks.
> >>>>>>
> >>>>>
> >>>>> We have the sd cache flush.  We would presumably just need to call blk_queue_write_cache()
> >>>>> for the !mmc_card_mmc(card) case e.g.
> >>>>>
> >>>>>           if (mmc_has_reliable_write(card)) {
> >>>>>                   md->flags |= MMC_BLK_REL_WR;
> >>>>>                   enable_fua = true;
> >>>>>           }
> >>>>>
> >>>>>           if (mmc_cache_enabled(card->host))
> >>>>>                   enable_cache = true;
> >>>>>
> >>>>>           blk_queue_write_cache(md->queue.queue, enable_cache, enable_fua);
> >>>>
> >>>> To me, this seems like the most reasonable thing to do.
> >>>>
> >>>> However, I have to admit that it's not clear to me, if there was a
> >>>> good reason to why commit f4c5522b0a88 ("mmc: Reliable write
> >>>> support.") also added support for REQ_FLUSH (write back cache) and why
> >>>> not only REQ_FUA. I assumed this was wrong too, right?
> >>>>
> >>
> >> Hi Ulf,
> >>
> >> 1. I've found the reason. If we only enable REQ_FUA, there won't be any
> >> effect -- The block layer won't send any request with FUA flag to the
> >> driver.
> >> If we want REQ_FUA to take effect, we must enable REQ_FLUSH. But on the
> >> contrary, REQ_FLUSH does not rely on REQ_FUA.
> >> In the previous patch(commit f4c5522b0a88 ("mmc: Reliable write
> >> support.")), REQ_FLUSH was added to make REQ_FUA effective. I've done
> >> experiments to prove this.
> >
> > Thanks for doing the research and for confirming.
> >
> > Note that this is also pretty well documented in
> > Documentation/block/writeback_cache_control.rst.
>
> Thanks for reminding. I'm clear now.
>
> >
> >>
> >> 2. Why block layer requires REQ_FLUSH to make REQ_FUA effective? I did
> >> not find the reason. Does anyone know about this? Thank you.
> >
> > The REQ_FLUSH indicates that the storage device has a write back
> > cache, which also can be flushed in some device specific way.
> >
> > The REQ_FUA (Force Unit Access), tells that the data can be written to
> > the storage device, in a way that when the I/O request is completed,
> > the data is fully written to the device (the data must not be left in
> > the write back cache). In other words, REQ_FUA doesn't make sense
> > unless REQ_FLUSH is supported too.
> >
>
> Thank you for your answer.
>
> > $subject patch should also conform to this pattern.
>
> I'm not sure if I understood this in a right way... Did you mean I
> should modify the subject of this mail/patch?

No, I just meant that the code in the patch should conform to this.

If REQ_FUA is set, REQ_FLUSH must be set too.

>
> >
> > However, it's still questionable to me whether we want to support
> > REQ_FUA through the eMMC reliable write command - in case we also have
> > support for managing the eMMC cache separately. It looks to me that
> > the reason why we currently support REQ_FUA, is because back in the
> > days when there was only the eMMC reliable write command available, it
> > was simply the best we could do. But it was never really a good fit.
> >
> > I am starting to think that we may consider dropping REQ_FUA, if we
> > have the option to manage the eMMC cache separately - no matter
> > whether the eMMC reliable write command is supported or not. In this
> > case, REQ_FLUSH is sufficient and also a better match to what we
> > really can support.
>
> Hi Ulf,
> As to dropping REQ_FUA, I don't know if it is a good idea, but generally
> we are facing three possible situations:
>
> 1. If both cache and reliable-write are available, both REQ_FUA and
> REQ_FLUSH can be supported at the same time. In this case, with
> available cache, the behavior of reliable-write is to write eMMC while
> skipping cache, which is consistent with the current kernel's definition
> of REQ_FUA. What's more, most eMMCs now support both cache and
> reliable-write command.

Yes, this seems reasonable.


> 2. If only reliable-write is available, REQ_FUA should not be supported,
> which is consistent with the current standard in another way. But I
> don't think eMMCs that only support reliable-write can be easily found
> nowadays.

If we drop REQ_FUA for this case, I am worried that we might break use
cases for those older eMMC devices.

So, no, let's keep REQ_FUA and REQ_FLUSH if reliable-write is supported.

> 3. If only cache is available, we just use REQ_FLUSH. It is not in
> conflict with keeping REQ_FUA.

Right.

>
> Maybe, is it more reasonable to reserve FUA and use if/else to pick it
> up or down, considering the compatibility? I mean, in most cases, FUA
> and FLUSH are complementary. So it seems more feasible with branch to
> choose.

Let's summarize what I think we should do then:

 if (reliable-write supported) {
     enable_fua = true;
     enable_cache = true;
}

if (mmc_cache_enabled)
     enable_cache = true;

blk_queue_write_cache(md->queue.queue, enable_cache, enable_fua);

Does this seem reasonable to you?

[...]

Kind regards
Uffe

Powered by blists - more mailing lists