On 3 February 2012 08:15, Jaehoon Chung <jh80.chung@xxxxxxxxxxx> wrote:
> On 02/03/2012 01:08 AM, S, Venkatraman wrote:
>
>> On Wed, Feb 1, 2012 at 8:57 PM, Saugata Das <saugata.das@xxxxxxxxxxxxxx> wrote:
>>> From: Saugata Das <saugata.das@xxxxxxxxxx>
>>>
>>> This patch groups the read or write transfers to eMMC in different contexts
>>> based on the block number. Transfers to consecutive blocks are grouped to a
>>> common context. So several small transfers combine to give performance like
>>> a large multi block transfer.
>>>
>>> The patch creates a context of 1MB multiple in non-large unit mode. Reliable
>>> mode is enabled in the context based on whether reliable write is enabled.
>>>
>>> Signed-off-by: Saugata Das <saugata.das@xxxxxxxxxx>
>>> ---
>> The patch subject could be better off as "mmc: eMMC4.5 context ID support"
>>
>>> drivers/mmc/card/block.c | 263 ++++++++++++++++++++++++++++++++++++++++++----
>>> drivers/mmc/core/mmc.c | 6 +
>>> include/linux/mmc/card.h | 13 +++
>>> include/linux/mmc/core.h | 9 ++
>>> include/linux/mmc/host.h | 1 +
>>> include/linux/mmc/mmc.h | 3 +
>>> 6 files changed, 275 insertions(+), 20 deletions(-)
>>>
>>> diff --git a/drivers/mmc/card/block.c b/drivers/mmc/card/block.c
>>> index 176b78e..161174a 100644
>>> --- a/drivers/mmc/card/block.c
>>> +++ b/drivers/mmc/card/block.c
>>> @@ -127,6 +127,8 @@ enum mmc_blk_status {
>>> module_param(perdev_minors, int, 0444);
>>> MODULE_PARM_DESC(perdev_minors, "Minors numbers to allocate per device");
>>>
>>> +static int mmc_blk_issue_rw_rq(struct mmc_queue *, struct request *);
>>> +
>>> static struct mmc_blk_data *mmc_blk_get(struct gendisk *disk)
>>> {
>>> struct mmc_blk_data *md;
>>> @@ -1071,6 +1073,192 @@ static int mmc_blk_err_check(struct mmc_card *card,
>>> return MMC_BLK_SUCCESS;
>>> }
>>>
>>> +/*
>>> + * Update the context information in the ext. CSD
>>
>> Mixed case
>>
>>> + */
>>> +static int mmc_context_modify(struct mmc_queue *mq,
>>> + struct mmc_card *card,
>>> + unsigned int context_id,
>>> + unsigned char context_conf)
>>> +{
>>> + /*
>>> + * Wait for any ongoing transfer
>>> + */
>>> + if (card->host->areq)
>>> + mmc_blk_issue_rw_rq(mq, NULL);
>>> +
>>> + /*
>>> + * CONTEXT_CONF array starts from context id 1
>>> + */
>>> + return mmc_switch(card, EXT_CSD_CMD_SET_NORMAL,
>>> + EXT_CSD_CONTEXT_CONF + context_id - 1,
>>> + context_conf,
>>> + card->ext_csd.generic_cmd6_time);
>>> +}
>>> +
>>> +/*
>>> + * Update timestamp, size and close context if needed
>>> + */
>>> +static int mmc_check_close_context(struct mmc_queue *mq,
>>> + struct mmc_card *card,
>>> + unsigned int context_id,
>>> + unsigned int status)
>>> +{
>>> + /*
>>> + * Check only for valid contexts
>>> + */
>>> + if ((context_id > 0) &&
>>> + (context_id <= card->ext_csd.max_context_id) &&
>>> + (card->mmc_context[context_id].valid)) {
>>> +
>>> + /*
>>> + * Incase of an error or we are multiple of erase block then
>>> + * close the context
>>> + */
>>> + if ((status) ||
>>> + ((card->mmc_context[context_id].size %
>>> + card->ext_csd.lu_size) == 0)) {
>>> + if (mmc_context_modify(mq, card, context_id,
>>> + MMC_CONTEXT_CLOSE))
>>> + return -1;
>> Should propagate the error code instead of returning -1
>>
>>> + card->mmc_context[context_id].valid = 0;
>>> + }
>>> + return 0;
>>> + }
>>> + return 0;
>>> +}
>>> +
>>> +/*
>>> + * Update timestamp, size and close context if needed
>>> + */
>>> +static int mmc_force_close_all_write_context(struct mmc_queue *mq,
>>> + struct mmc_card *card)
>>> +{
>>> + int i, ret = 0;
>>> + for (i = 1; i <= card->ext_csd.max_context_id; i++) {
>>> + if (card->mmc_context[i].direction != MMC_CONTEXT_READ) {
>>> + ret = mmc_check_close_context(mq, card, i,
>>> + MMC_CONTEXT_FORCE_CLOSE);
>>> + if (ret)
>>> + return ret;
>>> + }
>>> + }
>>> + return ret;
>>> +}
>>> +
>>> +/*
>>> + * Search for a context which is in the same direction and
>>> + * continuous in block numbers. If no matching context is
>>> + * found then take up an unused context. If all contexts are
>>> + * used then close the context which is least recently used,
>>> + * close it and the use it for the new transfer
>>> + */
>>> +static int mmc_get_context_id(struct mmc_queue *mq,
>>> + struct mmc_card *card,
>>> + unsigned int offset,
>>> + unsigned int size,
>>> + unsigned int direction,
>>> + unsigned int rel_wr)
>>> +{
>>> + int i, free_index = -1, lru_index = -1, ret_index;
>>> + unsigned int old_timestamp = -1;
>>> + unsigned int reliability_mode = rel_wr ? (0x01<<5) : 0;
>>> + struct mmc_context *context_ptr = &card->mmc_context[1];
>>> +
>>> + /*
>>> + * For older than 4.5 eMMC, there is no context ID
>>> + */
>>> + if (card->ext_csd.max_context_id == 0)
>>> + return 0;
>>> +
>>> + if (card->host->caps2 & MMC_CAP2_NO_CONTEXT)
>>> + return 0;
>>> +
>>> + /*
>>> + * If the request is LU size multiple then avoid
>>> + * using any context
>>> + */
>>> + if ((size % card->ext_csd.lu_size) == 0)
>>> + return 0;
>>> +
>>> + /*
>>> + * Context 0 is unused
>>> + */
>>> + for (i = 1; i <= card->ext_csd.max_context_id; i++, context_ptr++) {
>>> +
>>> + unsigned int context_start_blk;
>>> +
>>> + if (!context_ptr->valid) {
>>> + if (free_index == -1)
>>> + free_index = i;
>>> + continue;
>>> + }
>>> +
>>> + context_start_blk = context_ptr->offset -
>>> + context_ptr->size;
>>> +
>>> + if ((context_start_blk <= offset) &&
>>> + (context_ptr->offset > offset)) {
>>
>> How would this condition ever match ? The starting offset of the
>> previous request is usually less than the current request or it is
>> more, but not both.
>>
>>> + int ret = mmc_check_close_context(mq, card, i,
>>> + MMC_CONTEXT_FORCE_CLOSE);
>>> + if (ret < 0)
>>> + return ret;
>>> + if (free_index == -1)
>>> + free_index = i;
>>> + break;
>>> + }
>>> +
>>> + if ((lru_index == -1) ||
>>> + (context_ptr->timestamp < old_timestamp)) {
>>> + old_timestamp = context_ptr->timestamp;
>>> + lru_index = i;
>>> + }
>>> +
>>> + if ((context_ptr->direction != direction) ||
>>> + (context_ptr->offset != offset) ||
>>> + (context_ptr->reliability_mode !=
>>> + reliability_mode))
>>> + continue;
>>
>> I think the reliability mode match shouldn't matter. If the writes to
>> the same section of the device, the best of the reliability modes can
>> be applied.
>> This section should be moved above the context_start_blk and offset
>> checking code.
>> Why close an ongoing context if the directions don't match eventually ?
>>
>>> +
>>> + context_ptr->timestamp = jiffies;
>>> + context_ptr->offset += size;
>>> + context_ptr->size += size;
>>> + return i;
>>> + }
>>> +
>>> + if (free_index != -1) {
>>> + /*
>>> + * Open a free context
>>> + */
>>> + if (mmc_context_modify(mq, card, free_index,
>>> + (direction & 0x03) | reliability_mode))
>>> + return -1;
>>> + ret_index = free_index;
>>> + } else if (lru_index != -1) {
>>> + /*
>>> + * Close and reopen the LRU context
>>> + */
>>> + if (mmc_context_modify(mq, card, lru_index, MMC_CONTEXT_CLOSE))
>>> + return -1;
>>> +
>>> + if (mmc_context_modify(mq, card, lru_index,
>>> + (direction & 0x03) | reliability_mode))
>>> + return -1;
>>> + ret_index = lru_index;
>>> + } else
>>> + return -1;
>>> +
>>> + context_ptr = &card->mmc_context[ret_index];
>>> + context_ptr->offset = offset+size;
>>> + context_ptr->size = size;
>>> + context_ptr->direction = direction;
>>> + context_ptr->timestamp = jiffies;
>>> + context_ptr->reliability_mode = reliability_mode;
>>> + context_ptr->valid = 1;
>>> +
>>> + return ret_index;
>>> +}
>>> +
>>> static void mmc_blk_rw_rq_prep(struct mmc_queue_req *mqrq,
>>> struct mmc_card *card,
>>> int disable_multi,
>>> @@ -1080,7 +1268,6 @@ static void mmc_blk_rw_rq_prep(struct mmc_queue_req *mqrq,
>>> struct mmc_blk_request *brq = &mqrq->brq;
>>> struct request *req = mqrq->req;
>>> struct mmc_blk_data *md = mq->data;
>>> - bool do_data_tag;
>>>
>>> /*
>>> * Reliable writes are used to implement Forced Unit Access and
>>> @@ -1157,16 +1344,6 @@ static void mmc_blk_rw_rq_prep(struct mmc_queue_req *mqrq,
>>> mmc_apply_rel_rw(brq, card, req);
>>>
>>> /*
>>> - * Data tag is used only during writing meta data to speed
>>> - * up write and any subsequent read of this meta data
>>> - */
>>> - do_data_tag = (card->ext_csd.data_tag_unit_size) &&
>>> - (req->cmd_flags & REQ_META) &&
>>> - (rq_data_dir(req) == WRITE) &&
>>> - ((brq->data.blocks * brq->data.blksz) >=
>>> - card->ext_csd.data_tag_unit_size);
>>> -
>>> - /*
>>> * Pre-defined multi-block transfers are preferable to
>>> * open ended-ones (and necessary for reliable writes).
>>> * However, it is not sufficient to just send CMD23,
>>> @@ -1184,15 +1361,54 @@ static void mmc_blk_rw_rq_prep(struct mmc_queue_req *mqrq,
>>> * We'll avoid using CMD23-bounded multiblock writes for
>>> * these, while retaining features like reliable writes.
>>> */
>>> - if ((md->flags & MMC_BLK_CMD23) && mmc_op_multi(brq->cmd.opcode) &&
>>> - (do_rel_wr || !(card->quirks & MMC_QUIRK_BLK_NO_CMD23) ||
>>> - do_data_tag)) {
>>> - brq->sbc.opcode = MMC_SET_BLOCK_COUNT;
>>> - brq->sbc.arg = brq->data.blocks |
>>> - (do_rel_wr ? (1 << 31) : 0) |
>>> - (do_data_tag ? (1 << 29) : 0);
>>> - brq->sbc.flags = MMC_RSP_R1 | MMC_CMD_AC;
>>> - brq->mrq.sbc = &brq->sbc;
>>> + if ((md->flags & MMC_BLK_CMD23) && mmc_op_multi(brq->cmd.opcode)) {
>>> + /*
>>> + * Context ID is used for non-large unit and non-reliable
>>> + * write transfers provided the start block number
>>> + * sequentially follows any of the open contexts
>>> + */
>>> + int context_id;
>>> +
>>> + /*
>>> + * Data tag is used only during writing meta data to speed
>>> + * up write and any subsequent read of this meta data
>>> + */
>>> + bool do_data_tag = (card->ext_csd.data_tag_unit_size) &&
>>> + (req->cmd_flags & REQ_META) &&
>>> + (rq_data_dir(req) == WRITE) &&
>>> + ((brq->data.blocks * brq->data.blksz) >=
>>> + card->ext_csd.data_tag_unit_size);
>>> +
>>> + /*
>>> + * During fsync, ensure that data is committed to storage
>>> + */
>>> + if (req->cmd_flags & (REQ_FLUSH | REQ_SYNC))
>>> + mmc_force_close_all_write_context(mq, card);
>>> +
>>> + /*
>>> + * Context ID is used for non-large unit and non-reliable
>>> + * write transfers provided the start block number
>>> + * sequentially follows any of the open contexts
>>> + */
>>> + context_id = mmc_get_context_id(mq,
>>> + card,
>>> + blk_rq_pos(req),
>>> + brq->data.blocks,
>>> + (rq_data_dir(req) == WRITE) ? MMC_CONTEXT_WRITE :
>>> + MMC_CONTEXT_READ,
>>> + do_rel_wr);
>>> +
>>> + if (do_rel_wr || !(card->quirks & MMC_QUIRK_BLK_NO_CMD23) ||
>>> + do_data_tag || (context_id > 0)) {
>>> + brq->sbc.opcode = MMC_SET_BLOCK_COUNT;
>>> + brq->sbc.arg = brq->data.blocks |
>>> + (do_rel_wr ? (1 << 31) : 0) |
>>> + (do_data_tag ? (1 << 29) : 0) |
>>> + (((context_id > 0) && (!do_rel_wr)) ?
>>> + ((context_id & 0x0f) << 25) : 0);
>>> + brq->sbc.flags = MMC_RSP_R1 | MMC_CMD_AC;
>>> + brq->mrq.sbc = &brq->sbc;
>>> + }
>
> Can Data-tag & context-Id use together?
No. This is a mistake. I will rectify it in next version.
But overall, do we agree with this approach of context management
where we are trying to keep consecutive blocks within same context ?
>
> Best Regards,
> Jaehoon Chung
> --
> To unsubscribe from this list: send the line "unsubscribe linux-mmc" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-mmc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
