diff mbox

[v8] mmc: sdhci: Implement an SDHCI-specific bounce buffer

Message ID 20180128234453.2880-1-linus.walleij@linaro.org (mailing list archive)
State New, archived
Headers show

Commit Message

Linus Walleij Jan. 28, 2018, 11:44 p.m. UTC
The bounce buffer is gone from the MMC core, and now we found out
that there are some (crippled) i.MX boards out there that have broken
ADMA (cannot do scatter-gather), and also broken PIO so they must
use SDMA. Closer examination shows a less significant slowdown
also on SDMA-only capable Laptop hosts.

SDMA sets down the number of segments to one, so that each segment
gets turned into a singular request that ping-pongs to the block
layer before the next request/segment is issued.

Apparently it happens a lot that the block layer send requests
that include a lot of physically discontiguous segments. My guess
is that this phenomenon is coming from the file system.

These devices that cannot handle scatterlists in hardware can see
major benefits from a DMA-contiguous bounce buffer.

This patch accumulates those fragmented scatterlists in a physically
contiguous bounce buffer so that we can issue bigger DMA data chunks
to/from the card.

When tested with a PCI-integrated host (1217:8221) that
only supports SDMA:
0b:00.0 SD Host controller: O2 Micro, Inc. OZ600FJ0/OZ900FJ0/OZ600FJS
        SD/MMC Card Reader Controller (rev 05)
This patch gave ~1Mbyte/s improved throughput on large reads and
writes when testing using iozone than without the patch.

dmesg:
sdhci-pci 0000:0b:00.0: SDHCI controller found [1217:8221] (rev 5)
mmc0 bounce up to 128 segments into one, max segment size 65536 bytes
mmc0: SDHCI controller on PCI [0000:0b:00.0] using DMA

On the i.MX SDHCI controllers on the crippled i.MX 25 and i.MX 35
the patch restores the performance to what it was before we removed
the bounce buffers.

Cc: Pierre Ossman <pierre@ossman.eu>
Cc: Benoît Thébaudeau <benoit@wsystem.com>
Cc: Fabio Estevam <fabio.estevam@nxp.com>
Cc: Benjamin Beckmeyer <beckmeyer.b@rittal.de>
Cc: stable@vger.kernel.org # v4.14+
Fixes: de3ee99b097d ("mmc: Delete bounce buffer handling")
Tested-by: Benjamin Beckmeyer <beckmeyer.b@rittal.de>
Acked-by: Adrian Hunter <adrian.hunter@intel.com>
Signed-off-by: Linus Walleij <linus.walleij@linaro.org>
---
ChangeLog v7->v8:
- Fixed bad information and spelling mistakes in the commit
  message.
- Use sdhci_sdma_address() in one more spot identified by Adrian.
- Collected Adrian's ACK.
ChangeLog v6->v7:
- Fix the directions on dma_sync_for[device|cpu]() so the
  ownership of the buffer gets swapped properly and in the right
  direction for every transfer. Didn't see this because x86 PCI is
  DMA coherent...
- Tested and greelighted on i.MX 25.
- Also tested on the PCI version.
ChangeLog v5->v6:
- Again switch back to explicit sync of buffers. I want to get this
  solution to work because it gives more control and it's more
  elegant.
- Update host->max_req_size as noted by Adrian, hopefully this
  fixes the i.MX. I was just lucky on my Intel laptop I guess:
  the block stack never requested anything bigger than 64KB and
  that was why it worked even if max_req_size was bigger than
  what would fit in the bounce buffer.
- Copy the number of bytes in the mmc_data instead of the number
  of bytes in the bounce buffer. For RX this is blksize * blocks
  and for TX this is bytes_xfered.
- Break out a sdhci_sdma_address() for getting the DMA address
  for either the raw sglist or the bounce buffer depending on
  configuration.
- Add some explicit bounds check for the data so that we do not
  attempt to copy more than the bounce buffer size even if the
  block layer is erroneously configured.
- Move allocation of bounce buffer out to its own function.
- Use pr_[info|err] throughout so all debug prints from the
  driver come out in the same manner and style.
- Use unsigned int for the bounce buffer size.
- Re-tested with iozone: we still get the same nice performance
  improvements.
- Request a text on i.MX (hi Benjamin)
ChangeLog v4->v5:
- Go back to dma_alloc_coherent() as this apparently works better.
- Keep the other changes, cap for 64KB, fall back to single segments.
- Requesting a test of this on i.MX. (Sorry Benjamin.)
ChangeLog v3->v4:
- Cap the bounce buffer to 64KB instead of the biggest segment
  as we experience diminishing returns with buffers > 64KB.
- Instead of using dma_alloc_coherent(), use good old devm_kmalloc()
  and issue dma_sync_single_for*() to explicitly switch
  ownership between CPU and the device. This way we exercise the
  cache better and may consume less CPU.
- Bail out with single segments if we cannot allocate a bounce
  buffer.
- Tested on the PCI SDHCI on my laptop: requesting a new test
  on i.MX from Benjamin. (Please!)
ChangeLog v2->v3:
- Rewrite the commit message a bit
- Add Benjamin's Tested-by
- Add Fixes and stable tags
ChangeLog v1->v2:
- Skip the remapping and fiddling with the buffer, instead use
  dma_alloc_coherent() and use a simple, coherent bounce buffer.
- Couple kernel messages to ->parent of the mmc_host as it relates
  to the hardware characteristics.
---
 drivers/mmc/host/sdhci.c | 164 ++++++++++++++++++++++++++++++++++++++++++++---
 drivers/mmc/host/sdhci.h |   3 +
 2 files changed, 159 insertions(+), 8 deletions(-)

Comments

Ulf Hansson Jan. 31, 2018, 10:31 a.m. UTC | #1
On 29 January 2018 at 00:44, Linus Walleij <linus.walleij@linaro.org> wrote:
> The bounce buffer is gone from the MMC core, and now we found out
> that there are some (crippled) i.MX boards out there that have broken
> ADMA (cannot do scatter-gather), and also broken PIO so they must
> use SDMA. Closer examination shows a less significant slowdown
> also on SDMA-only capable Laptop hosts.
>
> SDMA sets down the number of segments to one, so that each segment
> gets turned into a singular request that ping-pongs to the block
> layer before the next request/segment is issued.
>
> Apparently it happens a lot that the block layer send requests
> that include a lot of physically discontiguous segments. My guess
> is that this phenomenon is coming from the file system.
>
> These devices that cannot handle scatterlists in hardware can see
> major benefits from a DMA-contiguous bounce buffer.
>
> This patch accumulates those fragmented scatterlists in a physically
> contiguous bounce buffer so that we can issue bigger DMA data chunks
> to/from the card.
>
> When tested with a PCI-integrated host (1217:8221) that
> only supports SDMA:
> 0b:00.0 SD Host controller: O2 Micro, Inc. OZ600FJ0/OZ900FJ0/OZ600FJS
>         SD/MMC Card Reader Controller (rev 05)
> This patch gave ~1Mbyte/s improved throughput on large reads and
> writes when testing using iozone than without the patch.
>
> dmesg:
> sdhci-pci 0000:0b:00.0: SDHCI controller found [1217:8221] (rev 5)
> mmc0 bounce up to 128 segments into one, max segment size 65536 bytes
> mmc0: SDHCI controller on PCI [0000:0b:00.0] using DMA
>
> On the i.MX SDHCI controllers on the crippled i.MX 25 and i.MX 35
> the patch restores the performance to what it was before we removed
> the bounce buffers.
>
> Cc: Pierre Ossman <pierre@ossman.eu>
> Cc: Benoît Thébaudeau <benoit@wsystem.com>
> Cc: Fabio Estevam <fabio.estevam@nxp.com>
> Cc: Benjamin Beckmeyer <beckmeyer.b@rittal.de>
> Cc: stable@vger.kernel.org # v4.14+
> Fixes: de3ee99b097d ("mmc: Delete bounce buffer handling")
> Tested-by: Benjamin Beckmeyer <beckmeyer.b@rittal.de>
> Acked-by: Adrian Hunter <adrian.hunter@intel.com>
> Signed-off-by: Linus Walleij <linus.walleij@linaro.org>

Thanks, applied for fixes!

Kind regards
Uffe

> ---
> ChangeLog v7->v8:
> - Fixed bad information and spelling mistakes in the commit
>   message.
> - Use sdhci_sdma_address() in one more spot identified by Adrian.
> - Collected Adrian's ACK.
> ChangeLog v6->v7:
> - Fix the directions on dma_sync_for[device|cpu]() so the
>   ownership of the buffer gets swapped properly and in the right
>   direction for every transfer. Didn't see this because x86 PCI is
>   DMA coherent...
> - Tested and greelighted on i.MX 25.
> - Also tested on the PCI version.
> ChangeLog v5->v6:
> - Again switch back to explicit sync of buffers. I want to get this
>   solution to work because it gives more control and it's more
>   elegant.
> - Update host->max_req_size as noted by Adrian, hopefully this
>   fixes the i.MX. I was just lucky on my Intel laptop I guess:
>   the block stack never requested anything bigger than 64KB and
>   that was why it worked even if max_req_size was bigger than
>   what would fit in the bounce buffer.
> - Copy the number of bytes in the mmc_data instead of the number
>   of bytes in the bounce buffer. For RX this is blksize * blocks
>   and for TX this is bytes_xfered.
> - Break out a sdhci_sdma_address() for getting the DMA address
>   for either the raw sglist or the bounce buffer depending on
>   configuration.
> - Add some explicit bounds check for the data so that we do not
>   attempt to copy more than the bounce buffer size even if the
>   block layer is erroneously configured.
> - Move allocation of bounce buffer out to its own function.
> - Use pr_[info|err] throughout so all debug prints from the
>   driver come out in the same manner and style.
> - Use unsigned int for the bounce buffer size.
> - Re-tested with iozone: we still get the same nice performance
>   improvements.
> - Request a text on i.MX (hi Benjamin)
> ChangeLog v4->v5:
> - Go back to dma_alloc_coherent() as this apparently works better.
> - Keep the other changes, cap for 64KB, fall back to single segments.
> - Requesting a test of this on i.MX. (Sorry Benjamin.)
> ChangeLog v3->v4:
> - Cap the bounce buffer to 64KB instead of the biggest segment
>   as we experience diminishing returns with buffers > 64KB.
> - Instead of using dma_alloc_coherent(), use good old devm_kmalloc()
>   and issue dma_sync_single_for*() to explicitly switch
>   ownership between CPU and the device. This way we exercise the
>   cache better and may consume less CPU.
> - Bail out with single segments if we cannot allocate a bounce
>   buffer.
> - Tested on the PCI SDHCI on my laptop: requesting a new test
>   on i.MX from Benjamin. (Please!)
> ChangeLog v2->v3:
> - Rewrite the commit message a bit
> - Add Benjamin's Tested-by
> - Add Fixes and stable tags
> ChangeLog v1->v2:
> - Skip the remapping and fiddling with the buffer, instead use
>   dma_alloc_coherent() and use a simple, coherent bounce buffer.
> - Couple kernel messages to ->parent of the mmc_host as it relates
>   to the hardware characteristics.
> ---
>  drivers/mmc/host/sdhci.c | 164 ++++++++++++++++++++++++++++++++++++++++++++---
>  drivers/mmc/host/sdhci.h |   3 +
>  2 files changed, 159 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
> index e9290a3439d5..d24306b2b839 100644
> --- a/drivers/mmc/host/sdhci.c
> +++ b/drivers/mmc/host/sdhci.c
> @@ -21,6 +21,7 @@
>  #include <linux/dma-mapping.h>
>  #include <linux/slab.h>
>  #include <linux/scatterlist.h>
> +#include <linux/sizes.h>
>  #include <linux/swiotlb.h>
>  #include <linux/regulator/consumer.h>
>  #include <linux/pm_runtime.h>
> @@ -502,8 +503,35 @@ static int sdhci_pre_dma_transfer(struct sdhci_host *host,
>         if (data->host_cookie == COOKIE_PRE_MAPPED)
>                 return data->sg_count;
>
> -       sg_count = dma_map_sg(mmc_dev(host->mmc), data->sg, data->sg_len,
> -                             mmc_get_dma_dir(data));
> +       /* Bounce write requests to the bounce buffer */
> +       if (host->bounce_buffer) {
> +               unsigned int length = data->blksz * data->blocks;
> +
> +               if (length > host->bounce_buffer_size) {
> +                       pr_err("%s: asked for transfer of %u bytes exceeds bounce buffer %u bytes\n",
> +                              mmc_hostname(host->mmc), length,
> +                              host->bounce_buffer_size);
> +                       return -EIO;
> +               }
> +               if (mmc_get_dma_dir(data) == DMA_TO_DEVICE) {
> +                       /* Copy the data to the bounce buffer */
> +                       sg_copy_to_buffer(data->sg, data->sg_len,
> +                                         host->bounce_buffer,
> +                                         length);
> +               }
> +               /* Switch ownership to the DMA */
> +               dma_sync_single_for_device(host->mmc->parent,
> +                                          host->bounce_addr,
> +                                          host->bounce_buffer_size,
> +                                          mmc_get_dma_dir(data));
> +               /* Just a dummy value */
> +               sg_count = 1;
> +       } else {
> +               /* Just access the data directly from memory */
> +               sg_count = dma_map_sg(mmc_dev(host->mmc),
> +                                     data->sg, data->sg_len,
> +                                     mmc_get_dma_dir(data));
> +       }
>
>         if (sg_count == 0)
>                 return -ENOSPC;
> @@ -673,6 +701,14 @@ static void sdhci_adma_table_post(struct sdhci_host *host,
>         }
>  }
>
> +static u32 sdhci_sdma_address(struct sdhci_host *host)
> +{
> +       if (host->bounce_buffer)
> +               return host->bounce_addr;
> +       else
> +               return sg_dma_address(host->data->sg);
> +}
> +
>  static u8 sdhci_calc_timeout(struct sdhci_host *host, struct mmc_command *cmd)
>  {
>         u8 count;
> @@ -858,8 +894,8 @@ static void sdhci_prepare_data(struct sdhci_host *host, struct mmc_command *cmd)
>                                              SDHCI_ADMA_ADDRESS_HI);
>                 } else {
>                         WARN_ON(sg_cnt != 1);
> -                       sdhci_writel(host, sg_dma_address(data->sg),
> -                               SDHCI_DMA_ADDRESS);
> +                       sdhci_writel(host, sdhci_sdma_address(host),
> +                                    SDHCI_DMA_ADDRESS);
>                 }
>         }
>
> @@ -2248,7 +2284,12 @@ static void sdhci_pre_req(struct mmc_host *mmc, struct mmc_request *mrq)
>
>         mrq->data->host_cookie = COOKIE_UNMAPPED;
>
> -       if (host->flags & SDHCI_REQ_USE_DMA)
> +       /*
> +        * No pre-mapping in the pre hook if we're using the bounce buffer,
> +        * for that we would need two bounce buffers since one buffer is
> +        * in flight when this is getting called.
> +        */
> +       if (host->flags & SDHCI_REQ_USE_DMA && !host->bounce_buffer)
>                 sdhci_pre_dma_transfer(host, mrq->data, COOKIE_PRE_MAPPED);
>  }
>
> @@ -2352,8 +2393,45 @@ static bool sdhci_request_done(struct sdhci_host *host)
>                 struct mmc_data *data = mrq->data;
>
>                 if (data && data->host_cookie == COOKIE_MAPPED) {
> -                       dma_unmap_sg(mmc_dev(host->mmc), data->sg, data->sg_len,
> -                                    mmc_get_dma_dir(data));
> +                       if (host->bounce_buffer) {
> +                               /*
> +                                * On reads, copy the bounced data into the
> +                                * sglist
> +                                */
> +                               if (mmc_get_dma_dir(data) == DMA_FROM_DEVICE) {
> +                                       unsigned int length = data->bytes_xfered;
> +
> +                                       if (length > host->bounce_buffer_size) {
> +                                               pr_err("%s: bounce buffer is %u bytes but DMA claims to have transferred %u bytes\n",
> +                                                      mmc_hostname(host->mmc),
> +                                                      host->bounce_buffer_size,
> +                                                      data->bytes_xfered);
> +                                               /* Cap it down and continue */
> +                                               length = host->bounce_buffer_size;
> +                                       }
> +                                       dma_sync_single_for_cpu(
> +                                               host->mmc->parent,
> +                                               host->bounce_addr,
> +                                               host->bounce_buffer_size,
> +                                               DMA_FROM_DEVICE);
> +                                       sg_copy_from_buffer(data->sg,
> +                                               data->sg_len,
> +                                               host->bounce_buffer,
> +                                               length);
> +                               } else {
> +                                       /* No copying, just switch ownership */
> +                                       dma_sync_single_for_cpu(
> +                                               host->mmc->parent,
> +                                               host->bounce_addr,
> +                                               host->bounce_buffer_size,
> +                                               mmc_get_dma_dir(data));
> +                               }
> +                       } else {
> +                               /* Unmap the raw data */
> +                               dma_unmap_sg(mmc_dev(host->mmc), data->sg,
> +                                            data->sg_len,
> +                                            mmc_get_dma_dir(data));
> +                       }
>                         data->host_cookie = COOKIE_UNMAPPED;
>                 }
>         }
> @@ -2636,7 +2714,8 @@ static void sdhci_data_irq(struct sdhci_host *host, u32 intmask)
>                  */
>                 if (intmask & SDHCI_INT_DMA_END) {
>                         u32 dmastart, dmanow;
> -                       dmastart = sg_dma_address(host->data->sg);
> +
> +                       dmastart = sdhci_sdma_address(host);
>                         dmanow = dmastart + host->data->bytes_xfered;
>                         /*
>                          * Force update to the next DMA block boundary.
> @@ -3217,6 +3296,68 @@ void __sdhci_read_caps(struct sdhci_host *host, u16 *ver, u32 *caps, u32 *caps1)
>  }
>  EXPORT_SYMBOL_GPL(__sdhci_read_caps);
>
> +static int sdhci_allocate_bounce_buffer(struct sdhci_host *host)
> +{
> +       struct mmc_host *mmc = host->mmc;
> +       unsigned int max_blocks;
> +       unsigned int bounce_size;
> +       int ret;
> +
> +       /*
> +        * Cap the bounce buffer at 64KB. Using a bigger bounce buffer
> +        * has diminishing returns, this is probably because SD/MMC
> +        * cards are usually optimized to handle this size of requests.
> +        */
> +       bounce_size = SZ_64K;
> +       /*
> +        * Adjust downwards to maximum request size if this is less
> +        * than our segment size, else hammer down the maximum
> +        * request size to the maximum buffer size.
> +        */
> +       if (mmc->max_req_size < bounce_size)
> +               bounce_size = mmc->max_req_size;
> +       max_blocks = bounce_size / 512;
> +
> +       /*
> +        * When we just support one segment, we can get significant
> +        * speedups by the help of a bounce buffer to group scattered
> +        * reads/writes together.
> +        */
> +       host->bounce_buffer = devm_kmalloc(mmc->parent,
> +                                          bounce_size,
> +                                          GFP_KERNEL);
> +       if (!host->bounce_buffer) {
> +               pr_err("%s: failed to allocate %u bytes for bounce buffer, falling back to single segments\n",
> +                      mmc_hostname(mmc),
> +                      bounce_size);
> +               /*
> +                * Exiting with zero here makes sure we proceed with
> +                * mmc->max_segs == 1.
> +                */
> +               return 0;
> +       }
> +
> +       host->bounce_addr = dma_map_single(mmc->parent,
> +                                          host->bounce_buffer,
> +                                          bounce_size,
> +                                          DMA_BIDIRECTIONAL);
> +       ret = dma_mapping_error(mmc->parent, host->bounce_addr);
> +       if (ret)
> +               /* Again fall back to max_segs == 1 */
> +               return 0;
> +       host->bounce_buffer_size = bounce_size;
> +
> +       /* Lie about this since we're bouncing */
> +       mmc->max_segs = max_blocks;
> +       mmc->max_seg_size = bounce_size;
> +       mmc->max_req_size = bounce_size;
> +
> +       pr_info("%s bounce up to %u segments into one, max segment size %u bytes\n",
> +               mmc_hostname(mmc), max_blocks, bounce_size);
> +
> +       return 0;
> +}
> +
>  int sdhci_setup_host(struct sdhci_host *host)
>  {
>         struct mmc_host *mmc;
> @@ -3713,6 +3854,13 @@ int sdhci_setup_host(struct sdhci_host *host)
>          */
>         mmc->max_blk_count = (host->quirks & SDHCI_QUIRK_NO_MULTIBLOCK) ? 1 : 65535;
>
> +       if (mmc->max_segs == 1) {
> +               /* This may alter mmc->*_blk_* parameters */
> +               ret = sdhci_allocate_bounce_buffer(host);
> +               if (ret)
> +                       return ret;
> +       }
> +
>         return 0;
>
>  unreg:
> diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
> index 54bc444c317f..1d7d61e25dbf 100644
> --- a/drivers/mmc/host/sdhci.h
> +++ b/drivers/mmc/host/sdhci.h
> @@ -440,6 +440,9 @@ struct sdhci_host {
>
>         int irq;                /* Device IRQ */
>         void __iomem *ioaddr;   /* Mapped address */
> +       char *bounce_buffer;    /* For packing SDMA reads/writes */
> +       dma_addr_t bounce_addr;
> +       unsigned int bounce_buffer_size;
>
>         const struct sdhci_ops *ops;    /* Low level hw interface */
>
> --
> 2.14.3
>
--
To unsubscribe from this list: send the line "unsubscribe linux-mmc" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
index e9290a3439d5..d24306b2b839 100644
--- a/drivers/mmc/host/sdhci.c
+++ b/drivers/mmc/host/sdhci.c
@@ -21,6 +21,7 @@ 
 #include <linux/dma-mapping.h>
 #include <linux/slab.h>
 #include <linux/scatterlist.h>
+#include <linux/sizes.h>
 #include <linux/swiotlb.h>
 #include <linux/regulator/consumer.h>
 #include <linux/pm_runtime.h>
@@ -502,8 +503,35 @@  static int sdhci_pre_dma_transfer(struct sdhci_host *host,
 	if (data->host_cookie == COOKIE_PRE_MAPPED)
 		return data->sg_count;
 
-	sg_count = dma_map_sg(mmc_dev(host->mmc), data->sg, data->sg_len,
-			      mmc_get_dma_dir(data));
+	/* Bounce write requests to the bounce buffer */
+	if (host->bounce_buffer) {
+		unsigned int length = data->blksz * data->blocks;
+
+		if (length > host->bounce_buffer_size) {
+			pr_err("%s: asked for transfer of %u bytes exceeds bounce buffer %u bytes\n",
+			       mmc_hostname(host->mmc), length,
+			       host->bounce_buffer_size);
+			return -EIO;
+		}
+		if (mmc_get_dma_dir(data) == DMA_TO_DEVICE) {
+			/* Copy the data to the bounce buffer */
+			sg_copy_to_buffer(data->sg, data->sg_len,
+					  host->bounce_buffer,
+					  length);
+		}
+		/* Switch ownership to the DMA */
+		dma_sync_single_for_device(host->mmc->parent,
+					   host->bounce_addr,
+					   host->bounce_buffer_size,
+					   mmc_get_dma_dir(data));
+		/* Just a dummy value */
+		sg_count = 1;
+	} else {
+		/* Just access the data directly from memory */
+		sg_count = dma_map_sg(mmc_dev(host->mmc),
+				      data->sg, data->sg_len,
+				      mmc_get_dma_dir(data));
+	}
 
 	if (sg_count == 0)
 		return -ENOSPC;
@@ -673,6 +701,14 @@  static void sdhci_adma_table_post(struct sdhci_host *host,
 	}
 }
 
+static u32 sdhci_sdma_address(struct sdhci_host *host)
+{
+	if (host->bounce_buffer)
+		return host->bounce_addr;
+	else
+		return sg_dma_address(host->data->sg);
+}
+
 static u8 sdhci_calc_timeout(struct sdhci_host *host, struct mmc_command *cmd)
 {
 	u8 count;
@@ -858,8 +894,8 @@  static void sdhci_prepare_data(struct sdhci_host *host, struct mmc_command *cmd)
 					     SDHCI_ADMA_ADDRESS_HI);
 		} else {
 			WARN_ON(sg_cnt != 1);
-			sdhci_writel(host, sg_dma_address(data->sg),
-				SDHCI_DMA_ADDRESS);
+			sdhci_writel(host, sdhci_sdma_address(host),
+				     SDHCI_DMA_ADDRESS);
 		}
 	}
 
@@ -2248,7 +2284,12 @@  static void sdhci_pre_req(struct mmc_host *mmc, struct mmc_request *mrq)
 
 	mrq->data->host_cookie = COOKIE_UNMAPPED;
 
-	if (host->flags & SDHCI_REQ_USE_DMA)
+	/*
+	 * No pre-mapping in the pre hook if we're using the bounce buffer,
+	 * for that we would need two bounce buffers since one buffer is
+	 * in flight when this is getting called.
+	 */
+	if (host->flags & SDHCI_REQ_USE_DMA && !host->bounce_buffer)
 		sdhci_pre_dma_transfer(host, mrq->data, COOKIE_PRE_MAPPED);
 }
 
@@ -2352,8 +2393,45 @@  static bool sdhci_request_done(struct sdhci_host *host)
 		struct mmc_data *data = mrq->data;
 
 		if (data && data->host_cookie == COOKIE_MAPPED) {
-			dma_unmap_sg(mmc_dev(host->mmc), data->sg, data->sg_len,
-				     mmc_get_dma_dir(data));
+			if (host->bounce_buffer) {
+				/*
+				 * On reads, copy the bounced data into the
+				 * sglist
+				 */
+				if (mmc_get_dma_dir(data) == DMA_FROM_DEVICE) {
+					unsigned int length = data->bytes_xfered;
+
+					if (length > host->bounce_buffer_size) {
+						pr_err("%s: bounce buffer is %u bytes but DMA claims to have transferred %u bytes\n",
+						       mmc_hostname(host->mmc),
+						       host->bounce_buffer_size,
+						       data->bytes_xfered);
+						/* Cap it down and continue */
+						length = host->bounce_buffer_size;
+					}
+					dma_sync_single_for_cpu(
+						host->mmc->parent,
+						host->bounce_addr,
+						host->bounce_buffer_size,
+						DMA_FROM_DEVICE);
+					sg_copy_from_buffer(data->sg,
+						data->sg_len,
+						host->bounce_buffer,
+						length);
+				} else {
+					/* No copying, just switch ownership */
+					dma_sync_single_for_cpu(
+						host->mmc->parent,
+						host->bounce_addr,
+						host->bounce_buffer_size,
+						mmc_get_dma_dir(data));
+				}
+			} else {
+				/* Unmap the raw data */
+				dma_unmap_sg(mmc_dev(host->mmc), data->sg,
+					     data->sg_len,
+					     mmc_get_dma_dir(data));
+			}
 			data->host_cookie = COOKIE_UNMAPPED;
 		}
 	}
@@ -2636,7 +2714,8 @@  static void sdhci_data_irq(struct sdhci_host *host, u32 intmask)
 		 */
 		if (intmask & SDHCI_INT_DMA_END) {
 			u32 dmastart, dmanow;
-			dmastart = sg_dma_address(host->data->sg);
+
+			dmastart = sdhci_sdma_address(host);
 			dmanow = dmastart + host->data->bytes_xfered;
 			/*
 			 * Force update to the next DMA block boundary.
@@ -3217,6 +3296,68 @@  void __sdhci_read_caps(struct sdhci_host *host, u16 *ver, u32 *caps, u32 *caps1)
 }
 EXPORT_SYMBOL_GPL(__sdhci_read_caps);
 
+static int sdhci_allocate_bounce_buffer(struct sdhci_host *host)
+{
+	struct mmc_host *mmc = host->mmc;
+	unsigned int max_blocks;
+	unsigned int bounce_size;
+	int ret;
+
+	/*
+	 * Cap the bounce buffer at 64KB. Using a bigger bounce buffer
+	 * has diminishing returns, this is probably because SD/MMC
+	 * cards are usually optimized to handle this size of requests.
+	 */
+	bounce_size = SZ_64K;
+	/*
+	 * Adjust downwards to maximum request size if this is less
+	 * than our segment size, else hammer down the maximum
+	 * request size to the maximum buffer size.
+	 */
+	if (mmc->max_req_size < bounce_size)
+		bounce_size = mmc->max_req_size;
+	max_blocks = bounce_size / 512;
+
+	/*
+	 * When we just support one segment, we can get significant
+	 * speedups by the help of a bounce buffer to group scattered
+	 * reads/writes together.
+	 */
+	host->bounce_buffer = devm_kmalloc(mmc->parent,
+					   bounce_size,
+					   GFP_KERNEL);
+	if (!host->bounce_buffer) {
+		pr_err("%s: failed to allocate %u bytes for bounce buffer, falling back to single segments\n",
+		       mmc_hostname(mmc),
+		       bounce_size);
+		/*
+		 * Exiting with zero here makes sure we proceed with
+		 * mmc->max_segs == 1.
+		 */
+		return 0;
+	}
+
+	host->bounce_addr = dma_map_single(mmc->parent,
+					   host->bounce_buffer,
+					   bounce_size,
+					   DMA_BIDIRECTIONAL);
+	ret = dma_mapping_error(mmc->parent, host->bounce_addr);
+	if (ret)
+		/* Again fall back to max_segs == 1 */
+		return 0;
+	host->bounce_buffer_size = bounce_size;
+
+	/* Lie about this since we're bouncing */
+	mmc->max_segs = max_blocks;
+	mmc->max_seg_size = bounce_size;
+	mmc->max_req_size = bounce_size;
+
+	pr_info("%s bounce up to %u segments into one, max segment size %u bytes\n",
+		mmc_hostname(mmc), max_blocks, bounce_size);
+
+	return 0;
+}
+
 int sdhci_setup_host(struct sdhci_host *host)
 {
 	struct mmc_host *mmc;
@@ -3713,6 +3854,13 @@  int sdhci_setup_host(struct sdhci_host *host)
 	 */
 	mmc->max_blk_count = (host->quirks & SDHCI_QUIRK_NO_MULTIBLOCK) ? 1 : 65535;
 
+	if (mmc->max_segs == 1) {
+		/* This may alter mmc->*_blk_* parameters */
+		ret = sdhci_allocate_bounce_buffer(host);
+		if (ret)
+			return ret;
+	}
+
 	return 0;
 
 unreg:
diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
index 54bc444c317f..1d7d61e25dbf 100644
--- a/drivers/mmc/host/sdhci.h
+++ b/drivers/mmc/host/sdhci.h
@@ -440,6 +440,9 @@  struct sdhci_host {
 
 	int irq;		/* Device IRQ */
 	void __iomem *ioaddr;	/* Mapped address */
+	char *bounce_buffer;	/* For packing SDMA reads/writes */
+	dma_addr_t bounce_addr;
+	unsigned int bounce_buffer_size;
 
 	const struct sdhci_ops *ops;	/* Low level hw interface */