Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:18 as permitted sender) client-ip=2620:137:e000::1:18;
MIME-Version: 1.0
References: <20220317111944.116148-1-yann.gautier@foss.st.com>
 <20220317111944.116148-3-yann.gautier@foss.st.com> <CAPDyKFqzzKgLHWiy26QW0hvM9kZEATS_c2mXkTuGiFpPaW8YKw@mail.gmail.com>
 <668661ca-271b-9a4f-6482-62f1b0190bb2@foss.st.com>
In-Reply-To: <668661ca-271b-9a4f-6482-62f1b0190bb2@foss.st.com>
From:   Ulf Hansson <ulf.hansson@linaro.org>
Date:   Fri, 25 Mar 2022 14:43:21 +0100
Message-ID: <CAPDyKFr3YGd4mytykFMAtESMg9fsRaAKZo5CHBN4hrUa8ekE8g@mail.gmail.com>
Subject: Re: [PATCH 2/2] mmc: mmci: stm32: use a buffer for unaligned DMA requests
To:     Yann Gautier <yann.gautier@foss.st.com>
Cc:     Christophe Kerello <christophe.kerello@foss.st.com>,
        Ludovic Barre <ludovic.barre@foss.st.com>,
        Maxime Coquelin <mcoquelin.stm32@gmail.com>,
        Alexandre Torgue <alexandre.torgue@foss.st.com>,
        Philipp Zabel <p.zabel@pengutronix.de>,
        Linus Walleij <linus.walleij@linaro.org>,
        linux-mmc@vger.kernel.org,
        linux-stm32@st-md-mailman.stormreply.com,
        linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Thu, 24 Mar 2022 at 17:23, Yann Gautier <yann.gautier@foss.st.com> wrote:
>
> On 3/24/22 12:55, Ulf Hansson wrote:
> > On Thu, 17 Mar 2022 at 12:19, Yann Gautier <yann.gautier@foss.st.com> wrote:
> >>
> >> In SDIO mode, the sg list for requests can be unaligned with what the
> >> STM32 SDMMC internal DMA can support. In that case, instead of failing,
> >> use a temporary bounce buffer to copy from/to the sg list.
> >> This buffer is limited to 1MB. But for that we need to also limit
> >> max_req_size to 1MB. It has not shown any throughput penalties for
> >> SD-cards or eMMC.
> >>
> >> Signed-off-by: Yann Gautier <yann.gautier@foss.st.com>
> >> ---
> >>   drivers/mmc/host/mmci_stm32_sdmmc.c | 80 +++++++++++++++++++++++------
> >>   1 file changed, 63 insertions(+), 17 deletions(-)
> >>
> >> diff --git a/drivers/mmc/host/mmci_stm32_sdmmc.c b/drivers/mmc/host/mmci_stm32_sdmmc.c
> >> index 4566d7fc9055..a4414e32800f 100644
> >> --- a/drivers/mmc/host/mmci_stm32_sdmmc.c
> >> +++ b/drivers/mmc/host/mmci_stm32_sdmmc.c
> >> @@ -43,6 +43,9 @@ struct sdmmc_lli_desc {
> >>   struct sdmmc_idma {
> >>          dma_addr_t sg_dma;
> >>          void *sg_cpu;
> >> +       dma_addr_t bounce_dma_addr;
> >> +       void *bounce_buf;
> >> +       bool use_bounce_buffer;
> >>   };
> >>
> >>   struct sdmmc_dlyb {
> >> @@ -54,6 +57,7 @@ struct sdmmc_dlyb {
> >>   static int sdmmc_idma_validate_data(struct mmci_host *host,
> >>                                      struct mmc_data *data)
> >>   {
> >> +       struct sdmmc_idma *idma = host->dma_priv;
> >>          struct scatterlist *sg;
> >>          int i;
> >>
> >> @@ -61,21 +65,23 @@ static int sdmmc_idma_validate_data(struct mmci_host *host,
> >>           * idma has constraints on idmabase & idmasize for each element
> >>           * excepted the last element which has no constraint on idmasize
> >>           */
> >> +       idma->use_bounce_buffer = false;
> >>          for_each_sg(data->sg, sg, data->sg_len - 1, i) {
> >>                  if (!IS_ALIGNED(sg->offset, sizeof(u32)) ||
> >>                      !IS_ALIGNED(sg->length, SDMMC_IDMA_BURST)) {
> >> -                       dev_err(mmc_dev(host->mmc),
> >> +                       dev_dbg(mmc_dev(host->mmc),
> >>                                  "unaligned scatterlist: ofst:%x length:%d\n",
> >>                                  data->sg->offset, data->sg->length);
> >> -                       return -EINVAL;
> >> +                       idma->use_bounce_buffer = true;
> >> +                       return 0;
> >>                  }
> >>          }
> >>
> >>          if (!IS_ALIGNED(sg->offset, sizeof(u32))) {
> >> -               dev_err(mmc_dev(host->mmc),
> >> +               dev_dbg(mmc_dev(host->mmc),
> >>                          "unaligned last scatterlist: ofst:%x length:%d\n",
> >>                          data->sg->offset, data->sg->length);
> >> -               return -EINVAL;
> >> +               idma->use_bounce_buffer = true;
> >>          }
> >>
> >>          return 0;
> >> @@ -84,18 +90,29 @@ static int sdmmc_idma_validate_data(struct mmci_host *host,
> >>   static int _sdmmc_idma_prep_data(struct mmci_host *host,
> >>                                   struct mmc_data *data)
> >>   {
> >> -       int n_elem;
> >> +       struct sdmmc_idma *idma = host->dma_priv;
> >>
> >> -       n_elem = dma_map_sg(mmc_dev(host->mmc),
> >> -                           data->sg,
> >> -                           data->sg_len,
> >> -                           mmc_get_dma_dir(data));
> >> +       if (idma->use_bounce_buffer) {
> >> +               if (data->flags & MMC_DATA_WRITE) {
> >> +                       unsigned int xfer_bytes = data->blksz * data->blocks;
> >>
> >> -       if (!n_elem) {
> >> -               dev_err(mmc_dev(host->mmc), "dma_map_sg failed\n");
> >> -               return -EINVAL;
> >> -       }
> >> +                       sg_copy_to_buffer(data->sg, data->sg_len,
> >> +                                         idma->bounce_buf, xfer_bytes);
> >> +                       dma_wmb();
> >> +               }
> >> +       } else {
> >> +               int n_elem;
> >> +
> >> +               n_elem = dma_map_sg(mmc_dev(host->mmc),
> >> +                                   data->sg,
> >> +                                   data->sg_len,
> >> +                                   mmc_get_dma_dir(data));
> >>
> >> +               if (!n_elem) {
> >> +                       dev_err(mmc_dev(host->mmc), "dma_map_sg failed\n");
> >> +                       return -EINVAL;
> >> +               }
> >> +       }
> >>          return 0;
> >>   }
> >>
> >> @@ -112,8 +129,19 @@ static int sdmmc_idma_prep_data(struct mmci_host *host,
> >>   static void sdmmc_idma_unprep_data(struct mmci_host *host,
> >>                                     struct mmc_data *data, int err)
> >>   {
> >> -       dma_unmap_sg(mmc_dev(host->mmc), data->sg, data->sg_len,
> >> -                    mmc_get_dma_dir(data));
> >> +       struct sdmmc_idma *idma = host->dma_priv;
> >> +
> >> +       if (idma->use_bounce_buffer) {
> >> +               if (data->flags & MMC_DATA_READ) {
> >> +                       unsigned int xfer_bytes = data->blksz * data->blocks;
> >> +
> >> +                       sg_copy_from_buffer(data->sg, data->sg_len,
> >> +                                           idma->bounce_buf, xfer_bytes);
> >> +               }
> >> +       } else {
> >> +               dma_unmap_sg(mmc_dev(host->mmc), data->sg, data->sg_len,
> >> +                            mmc_get_dma_dir(data));
> >> +       }
> >>   }
> >>
> >>   static int sdmmc_idma_setup(struct mmci_host *host)
> >> @@ -137,6 +165,16 @@ static int sdmmc_idma_setup(struct mmci_host *host)
> >>                  host->mmc->max_segs = SDMMC_LLI_BUF_LEN /
> >>                          sizeof(struct sdmmc_lli_desc);
> >>                  host->mmc->max_seg_size = host->variant->stm32_idmabsize_mask;
> >> +
> >> +               host->mmc->max_req_size = SZ_1M;
> >> +               idma->bounce_buf = dmam_alloc_coherent(dev,
> >> +                                                      host->mmc->max_req_size,
> >> +                                                      &idma->bounce_dma_addr,
> >> +                                                      GFP_KERNEL);
> >> +               if (!idma->bounce_buf) {
> >> +                       dev_err(dev, "Unable to map allocate DMA bounce buffer.\n");
> >> +                       return -ENOMEM;
> >
> Hi Ulf,
>
> > If we fail to allocate the 1M bounce buffer, then we end up always
> > using a PIO based mode, right?
> >
> > Perhaps we can allow the above allocation to fail, but then limit us
> > to use DMA only when the buffers are properly aligned? Would that
> > work?
> >
> We have never supported PIO mode with STM32 variant.
> We only support DMA single buffer or DMA LLI.
> As we cannot have DMA LLI for unaligned access, we'll default to single
> mode.

Right, I was looking at the legacy variant, which uses PIO as
fallback. Sorry for my ignorance.

> If allocation fails, it then won't work.

Right, but that's only part of the issue, I think.

> Maybe we shouldn't fail here, and just check idma->bounce_buf in
> validate data function. If buffer is not allocated, we just return
> -EINVAL as it was done before.

Yes, something along those lines. However, there is another problem
too, which is that the allocation will be done for each instance of
the host that is probed. In all cases but the SDIO case, this would be
a waste, right?

Perhaps we should manage the allocation in the validate function too
(de-allocation should be handled at ->remove()). In this way, the
buffer will only be allocated when it's actually needed. Yes, it would
add a latency while serving the *first* request that has unaligned
buffers, but I guess we can live with that?

>
> Best regards,
> Yann

Kind regards
Uffe

>
> >> +               }
> >>          } else {
> >>                  host->mmc->max_segs = 1;
> >>                  host->mmc->max_seg_size = host->mmc->max_req_size;
> >> @@ -154,8 +192,16 @@ static int sdmmc_idma_start(struct mmci_host *host, unsigned int *datactrl)
> >>          struct scatterlist *sg;
> >>          int i;
> >>
> >> -       if (!host->variant->dma_lli || data->sg_len == 1) {
> >> -               writel_relaxed(sg_dma_address(data->sg),
> >> +       if (!host->variant->dma_lli || data->sg_len == 1 ||
> >> +           idma->use_bounce_buffer) {
> >> +               u32 dma_addr;
> >> +
> >> +               if (idma->use_bounce_buffer)
> >> +                       dma_addr = idma->bounce_dma_addr;
> >> +               else
> >> +                       dma_addr = sg_dma_address(data->sg);
> >> +
> >> +               writel_relaxed(dma_addr,
> >>                                 host->base + MMCI_STM32_IDMABASE0R);
> >>                  writel_relaxed(MMCI_STM32_IDMAEN,
> >>                                 host->base + MMCI_STM32_IDMACTRLR);
> >
> > Kind regards
> > Uffe
>