It is expected that Extended CSD register(the size of this register
is larger than CID/CSD) will be referenced more frequently as more
fields have been added to Extended CSD and it seems that it is not
a good option to double the memory used.
This patch is intended to avoid the use of bounce buffer for reading
Extended CSD register in mmc_send_cxd_data(). It will provide a better
performance gain by removing memcpy() overhead for a half KiB and
a redundant bounce buffer allocated repeatedly at the cost of providing
DMA-capable buffer from upper caller(but on-stack buffer is allowed
with no performance gain).
Signed-off-by: Kyungsik Lee <[email protected]>
Signed-off-by: S, Venkatraman <[email protected]>
---
Changes in v2:
- Handling on-stack buffer if it's used in caller.
Changes in v3:
- Remove unnecesary code.
Changes in v4:
- Modify codes based-on S, Venkatraman's comments.
Changes in v5:
- Modify commit log and add NOTE description.
---
drivers/mmc/core/mmc_ops.c | 58 +++++++++++++++++++++++++++++++------------
1 files changed, 42 insertions(+), 16 deletions(-)
diff --git a/drivers/mmc/core/mmc_ops.c b/drivers/mmc/core/mmc_ops.c
index 0ed2cc5..acf578d 100644
--- a/drivers/mmc/core/mmc_ops.c
+++ b/drivers/mmc/core/mmc_ops.c
@@ -230,6 +230,10 @@ mmc_send_cxd_native(struct mmc_host *host, u32 arg, u32 *cxd, int opcode)
return 0;
}
+/*
+ * NOTE: void *buf, Caller for the buf is required to use DMA-capable
+ * buffer or on-stack buffer(with some overhead in callee).
+ */
static int
mmc_send_cxd_data(struct mmc_card *card, struct mmc_host *host,
u32 opcode, void *buf, unsigned len)
@@ -239,13 +243,19 @@ mmc_send_cxd_data(struct mmc_card *card, struct mmc_host *host,
struct mmc_data data = {0};
struct scatterlist sg;
void *data_buf;
+ int is_on_stack;
- /* dma onto stack is unsafe/nonportable, but callers to this
- * routine normally provide temporary on-stack buffers ...
- */
- data_buf = kmalloc(len, GFP_KERNEL);
- if (data_buf == NULL)
- return -ENOMEM;
+ is_on_stack = object_is_on_stack(buf);
+ if (is_on_stack) {
+
+ /* dma onto stack is unsafe/nonportable, but callers to this
+ * routine normally provide temporary on-stack buffers ...
+ */
+ data_buf = kmalloc(len, GFP_KERNEL);
+ if (data_buf == NULL)
+ return -ENOMEM;
+ } else
+ data_buf = buf;
mrq.cmd = &cmd;
mrq.data = &data;
@@ -280,8 +290,10 @@ mmc_send_cxd_data(struct mmc_card *card, struct mmc_host *host,
mmc_wait_for_req(host, &mrq);
- memcpy(buf, data_buf, len);
- kfree(data_buf);
+ if (is_on_stack) {
+ memcpy(buf, data_buf, len);
+ kfree(data_buf);
+ }
if (cmd.error)
return cmd.error;
@@ -294,24 +306,32 @@ mmc_send_cxd_data(struct mmc_card *card, struct mmc_host *host,
int mmc_send_csd(struct mmc_card *card, u32 *csd)
{
int ret, i;
+ u32 *csd_tmp;
if (!mmc_host_is_spi(card->host))
return mmc_send_cxd_native(card->host, card->rca << 16,
csd, MMC_SEND_CSD);
- ret = mmc_send_cxd_data(card, card->host, MMC_SEND_CSD, csd, 16);
+ csd_tmp = kmalloc(16, GFP_KERNEL);
+ if (!csd_tmp)
+ return -ENOMEM;
+
+ ret = mmc_send_cxd_data(card, card->host, MMC_SEND_CSD, csd_tmp, 16);
if (ret)
- return ret;
+ goto err;
for (i = 0;i < 4;i++)
- csd[i] = be32_to_cpu(csd[i]);
+ csd[i] = be32_to_cpu(csd_tmp[i]);
- return 0;
+err:
+ kfree(csd_tmp);
+ return ret;
}
int mmc_send_cid(struct mmc_host *host, u32 *cid)
{
int ret, i;
+ u32 *cid_tmp;
if (!mmc_host_is_spi(host)) {
if (!host->card)
@@ -320,14 +340,20 @@ int mmc_send_cid(struct mmc_host *host, u32 *cid)
cid, MMC_SEND_CID);
}
- ret = mmc_send_cxd_data(NULL, host, MMC_SEND_CID, cid, 16);
+ cid_tmp = kmalloc(16, GFP_KERNEL);
+ if (!cid_tmp)
+ return -ENOMEM;
+
+ ret = mmc_send_cxd_data(NULL, host, MMC_SEND_CID, cid_tmp, 16);
if (ret)
- return ret;
+ goto err;
for (i = 0;i < 4;i++)
- cid[i] = be32_to_cpu(cid[i]);
+ cid[i] = be32_to_cpu(cid_tmp[i]);
- return 0;
+err:
+ kfree(cid_tmp);
+ return ret;
}
int mmc_send_ext_csd(struct mmc_card *card, u8 *ext_csd)
--
1.7.0.4
On Thu, Aug 2, 2012 at 12:37 PM, Kyungsik Lee <[email protected]> wrote:
> It is expected that Extended CSD register(the size of this register
> is larger than CID/CSD) will be referenced more frequently as more
> fields have been added to Extended CSD and it seems that it is not
> a good option to double the memory used.
>
> This patch is intended to avoid the use of bounce buffer for reading
> Extended CSD register in mmc_send_cxd_data(). It will provide a better
> performance gain by removing memcpy() overhead for a half KiB and
> a redundant bounce buffer allocated repeatedly at the cost of providing
> DMA-capable buffer from upper caller(but on-stack buffer is allowed
> with no performance gain).
>
> Signed-off-by: Kyungsik Lee <[email protected]>
> Signed-off-by: S, Venkatraman <[email protected]>
I don't think my sign off is warranted. I gave just an illustration.
You can add my
Reviewed-by: Venkatraman S <[email protected]>
Regards,
Venkat.