mmc: fix async request mechanism for sequential read scenarios
authorKonstantin Dorfman <kdorfman@codeaurora.org>
Mon, 14 Jan 2013 19:28:17 +0000 (14:28 -0500)
committerChris Ball <cjb@laptop.org>
Mon, 28 Jan 2013 11:51:33 +0000 (06:51 -0500)
When current request is running on the bus and if next request fetched
by mmcqd is NULL, mmc context (mmcqd thread) gets blocked until the
current request completes. This means that if new request comes in while
the mmcqd thread is blocked, this new request can not be prepared in
parallel to current ongoing request. This may result in delaying the new
request execution and increase it's latency.

This change allows to wake up the MMC thread on new request arrival.
Now once the MMC thread is woken up, a new request can be fetched and
prepared in parallel to the current running request which means this new
request can be started immediately after the current running request
completes.

With this change read throughput is improved by 16%.

Signed-off-by: Konstantin Dorfman <kdorfman@codeaurora.org>
Reviewed-by: Seungwon Jeon <tgih.jun@samsung.com>
Signed-off-by: Chris Ball <cjb@laptop.org>
drivers/mmc/card/block.c
drivers/mmc/card/queue.c
drivers/mmc/card/queue.h
drivers/mmc/core/bus.c
drivers/mmc/core/core.c
drivers/mmc/core/core.h
include/linux/mmc/card.h
include/linux/mmc/core.h
include/linux/mmc/host.h

index 21056b9ef0a0f97bf9f1e71de1e78a3253cd5284..f79b4688e4718f83f792d31a3221f40fd6450edb 100644 (file)
@@ -113,17 +113,6 @@ struct mmc_blk_data {
 
 static DEFINE_MUTEX(open_lock);
 
-enum mmc_blk_status {
-       MMC_BLK_SUCCESS = 0,
-       MMC_BLK_PARTIAL,
-       MMC_BLK_CMD_ERR,
-       MMC_BLK_RETRY,
-       MMC_BLK_ABORT,
-       MMC_BLK_DATA_ERR,
-       MMC_BLK_ECC_ERR,
-       MMC_BLK_NOMEDIUM,
-};
-
 module_param(perdev_minors, int, 0444);
 MODULE_PARM_DESC(perdev_minors, "Minors numbers to allocate per device");
 
@@ -1364,8 +1353,11 @@ static int mmc_blk_issue_rw_rq(struct mmc_queue *mq, struct request *rqc)
                } else
                        areq = NULL;
                areq = mmc_start_req(card->host, areq, (int *) &status);
-               if (!areq)
+               if (!areq) {
+                       if (status == MMC_BLK_NEW_REQUEST)
+                               mq->flags |= MMC_QUEUE_NEW_REQUEST;
                        return 0;
+               }
 
                mq_rq = container_of(areq, struct mmc_queue_req, mmc_active);
                brq = &mq_rq->brq;
@@ -1438,6 +1430,10 @@ static int mmc_blk_issue_rw_rq(struct mmc_queue *mq, struct request *rqc)
                        break;
                case MMC_BLK_NOMEDIUM:
                        goto cmd_abort;
+               default:
+                       pr_err("%s: Unhandled return value (%d)",
+                                       req->rq_disk->disk_name, status);
+                       goto cmd_abort;
                }
 
                if (ret) {
@@ -1472,6 +1468,8 @@ static int mmc_blk_issue_rq(struct mmc_queue *mq, struct request *req)
        int ret;
        struct mmc_blk_data *md = mq->data;
        struct mmc_card *card = md->queue.card;
+       struct mmc_host *host = card->host;
+       unsigned long flags;
 
        if (req && !mq->mqrq_prev->req)
                /* claim host only for the first request */
@@ -1486,6 +1484,7 @@ static int mmc_blk_issue_rq(struct mmc_queue *mq, struct request *req)
                goto out;
        }
 
+       mq->flags &= ~MMC_QUEUE_NEW_REQUEST;
        if (req && req->cmd_flags & REQ_DISCARD) {
                /* complete ongoing async transfer before issuing discard */
                if (card->host->areq)
@@ -1501,11 +1500,16 @@ static int mmc_blk_issue_rq(struct mmc_queue *mq, struct request *req)
                        mmc_blk_issue_rw_rq(mq, NULL);
                ret = mmc_blk_issue_flush(mq, req);
        } else {
+               if (!req && host->areq) {
+                       spin_lock_irqsave(&host->context_info.lock, flags);
+                       host->context_info.is_waiting_last_req = true;
+                       spin_unlock_irqrestore(&host->context_info.lock, flags);
+               }
                ret = mmc_blk_issue_rw_rq(mq, req);
        }
 
 out:
-       if (!req)
+       if (!req && !(mq->flags & MMC_QUEUE_NEW_REQUEST))
                /* release host only when there are no more requests */
                mmc_release_host(card->host);
        return ret;
index d630d9861e7bf0f4631b99374b84b1b0baa180d4..5e0971016ac5fb0043f8bda840eebce4ab68310a 100644 (file)
@@ -22,7 +22,6 @@
 
 #define MMC_QUEUE_BOUNCESZ     65536
 
-#define MMC_QUEUE_SUSPENDED    (1 << 0)
 
 #define MMC_REQ_SPECIAL_MASK   (REQ_DISCARD | REQ_FLUSH)
 
@@ -72,6 +71,10 @@ static int mmc_queue_thread(void *d)
                        set_current_state(TASK_RUNNING);
                        cmd_flags = req ? req->cmd_flags : 0;
                        mq->issue_fn(mq, req);
+                       if (mq->flags & MMC_QUEUE_NEW_REQUEST) {
+                               mq->flags &= ~MMC_QUEUE_NEW_REQUEST;
+                               continue; /* fetch again */
+                       }
 
                        /*
                         * Current request becomes previous request
@@ -113,6 +116,8 @@ static void mmc_request_fn(struct request_queue *q)
 {
        struct mmc_queue *mq = q->queuedata;
        struct request *req;
+       unsigned long flags;
+       struct mmc_context_info *cntx;
 
        if (!mq) {
                while ((req = blk_fetch_request(q)) != NULL) {
@@ -122,7 +127,20 @@ static void mmc_request_fn(struct request_queue *q)
                return;
        }
 
-       if (!mq->mqrq_cur->req && !mq->mqrq_prev->req)
+       cntx = &mq->card->host->context_info;
+       if (!mq->mqrq_cur->req && mq->mqrq_prev->req) {
+               /*
+                * New MMC request arrived when MMC thread may be
+                * blocked on the previous request to be complete
+                * with no current request fetched
+                */
+               spin_lock_irqsave(&cntx->lock, flags);
+               if (cntx->is_waiting_last_req) {
+                       cntx->is_new_req = true;
+                       wake_up_interruptible(&cntx->wait);
+               }
+               spin_unlock_irqrestore(&cntx->lock, flags);
+       } else if (!mq->mqrq_cur->req && !mq->mqrq_prev->req)
                wake_up_process(mq->thread);
 }
 
index d2a1eb4b9f9fade36848726f4e36fc4db1463cca..e20c27b2b8b4814ff8c46497babd0c572f125dbf 100644 (file)
@@ -27,6 +27,9 @@ struct mmc_queue {
        struct task_struct      *thread;
        struct semaphore        thread_sem;
        unsigned int            flags;
+#define MMC_QUEUE_SUSPENDED    (1 << 0)
+#define MMC_QUEUE_NEW_REQUEST  (1 << 1)
+
        int                     (*issue_fn)(struct mmc_queue *, struct request *);
        void                    *data;
        struct request_queue    *queue;
index 420cb6753c1e74c48a604b27cd9fc54752390085..e219c97a02a416e2439f6640093ac29de4e5d371 100644 (file)
@@ -321,6 +321,7 @@ int mmc_add_card(struct mmc_card *card)
 #ifdef CONFIG_DEBUG_FS
        mmc_add_card_debugfs(card);
 #endif
+       mmc_init_context_info(card->host);
 
        ret = device_add(&card->dev);
        if (ret)
index aaed7687cf094fd9659dae2892b89c9ccd89c138..8b3a1222e665a8a8ac087bf2ced4ac66565add76 100644 (file)
@@ -319,11 +319,44 @@ out:
 }
 EXPORT_SYMBOL(mmc_start_bkops);
 
+/*
+ * mmc_wait_data_done() - done callback for data request
+ * @mrq: done data request
+ *
+ * Wakes up mmc context, passed as a callback to host controller driver
+ */
+static void mmc_wait_data_done(struct mmc_request *mrq)
+{
+       mrq->host->context_info.is_done_rcv = true;
+       wake_up_interruptible(&mrq->host->context_info.wait);
+}
+
 static void mmc_wait_done(struct mmc_request *mrq)
 {
        complete(&mrq->completion);
 }
 
+/*
+ *__mmc_start_data_req() - starts data request
+ * @host: MMC host to start the request
+ * @mrq: data request to start
+ *
+ * Sets the done callback to be called when request is completed by the card.
+ * Starts data mmc request execution
+ */
+static int __mmc_start_data_req(struct mmc_host *host, struct mmc_request *mrq)
+{
+       mrq->done = mmc_wait_data_done;
+       mrq->host = host;
+       if (mmc_card_removed(host->card)) {
+               mrq->cmd->error = -ENOMEDIUM;
+               return -ENOMEDIUM;
+       }
+       mmc_start_request(host, mrq);
+
+       return 0;
+}
+
 static int __mmc_start_req(struct mmc_host *host, struct mmc_request *mrq)
 {
        init_completion(&mrq->completion);
@@ -337,6 +370,62 @@ static int __mmc_start_req(struct mmc_host *host, struct mmc_request *mrq)
        return 0;
 }
 
+/*
+ * mmc_wait_for_data_req_done() - wait for request completed
+ * @host: MMC host to prepare the command.
+ * @mrq: MMC request to wait for
+ *
+ * Blocks MMC context till host controller will ack end of data request
+ * execution or new request notification arrives from the block layer.
+ * Handles command retries.
+ *
+ * Returns enum mmc_blk_status after checking errors.
+ */
+static int mmc_wait_for_data_req_done(struct mmc_host *host,
+                                     struct mmc_request *mrq,
+                                     struct mmc_async_req *next_req)
+{
+       struct mmc_command *cmd;
+       struct mmc_context_info *context_info = &host->context_info;
+       int err;
+       unsigned long flags;
+
+       while (1) {
+               wait_event_interruptible(context_info->wait,
+                               (context_info->is_done_rcv ||
+                                context_info->is_new_req));
+               spin_lock_irqsave(&context_info->lock, flags);
+               context_info->is_waiting_last_req = false;
+               spin_unlock_irqrestore(&context_info->lock, flags);
+               if (context_info->is_done_rcv) {
+                       context_info->is_done_rcv = false;
+                       context_info->is_new_req = false;
+                       cmd = mrq->cmd;
+                       if (!cmd->error || !cmd->retries ||
+                           mmc_card_removed(host->card)) {
+                               err = host->areq->err_check(host->card,
+                                                           host->areq);
+                               break; /* return err */
+                       } else {
+                               pr_info("%s: req failed (CMD%u): %d, retrying...\n",
+                                       mmc_hostname(host),
+                                       cmd->opcode, cmd->error);
+                               cmd->retries--;
+                               cmd->error = 0;
+                               host->ops->request(host, mrq);
+                               continue; /* wait for done/new event again */
+                       }
+               } else if (context_info->is_new_req) {
+                       context_info->is_new_req = false;
+                       if (!next_req) {
+                               err = MMC_BLK_NEW_REQUEST;
+                               break; /* return err */
+                       }
+               }
+       }
+       return err;
+}
+
 static void mmc_wait_for_req_done(struct mmc_host *host,
                                  struct mmc_request *mrq)
 {
@@ -426,8 +515,17 @@ struct mmc_async_req *mmc_start_req(struct mmc_host *host,
                mmc_pre_req(host, areq->mrq, !host->areq);
 
        if (host->areq) {
-               mmc_wait_for_req_done(host, host->areq->mrq);
-               err = host->areq->err_check(host->card, host->areq);
+                       err = mmc_wait_for_data_req_done(host, host->areq->mrq,
+                                       areq);
+                       if (err == MMC_BLK_NEW_REQUEST) {
+                               if (error)
+                                       *error = err;
+                               /*
+                                * The previous request was not completed,
+                                * nothing to return
+                                */
+                               return NULL;
+                       }
                /*
                 * Check BKOPS urgency for each R1 response
                 */
@@ -439,7 +537,7 @@ struct mmc_async_req *mmc_start_req(struct mmc_host *host,
        }
 
        if (!err && areq)
-               start_err = __mmc_start_req(host, areq->mrq);
+               start_err = __mmc_start_data_req(host, areq->mrq);
 
        if (host->areq)
                mmc_post_req(host, host->areq->mrq, 0);
@@ -2581,6 +2679,23 @@ int mmc_pm_notify(struct notifier_block *notify_block,
 }
 #endif
 
+/**
+ * mmc_init_context_info() - init synchronization context
+ * @host: mmc host
+ *
+ * Init struct context_info needed to implement asynchronous
+ * request mechanism, used by mmc core, host driver and mmc requests
+ * supplier.
+ */
+void mmc_init_context_info(struct mmc_host *host)
+{
+       spin_lock_init(&host->context_info.lock);
+       host->context_info.is_new_req = false;
+       host->context_info.is_done_rcv = false;
+       host->context_info.is_waiting_last_req = false;
+       init_waitqueue_head(&host->context_info.wait);
+}
+
 static int __init mmc_init(void)
 {
        int ret;
index 3bdafbca354f4ecdce7f4db56ce9baf82ba11cbb..0272b3284b5efcb4d659e90d32ec1d41fb2ce78a 100644 (file)
@@ -76,5 +76,6 @@ void mmc_remove_host_debugfs(struct mmc_host *host);
 void mmc_add_card_debugfs(struct mmc_card *card);
 void mmc_remove_card_debugfs(struct mmc_card *card);
 
+void mmc_init_context_info(struct mmc_host *host);
 #endif
 
index 5c69315d60cc6ac8a4a3b5b447e9be514ecb4123..be2500a499251bfc4c5b40512e61acbfbb5bae47 100644 (file)
@@ -187,6 +187,18 @@ struct sdio_func_tuple;
 
 #define SDIO_MAX_FUNCS         7
 
+enum mmc_blk_status {
+       MMC_BLK_SUCCESS = 0,
+       MMC_BLK_PARTIAL,
+       MMC_BLK_CMD_ERR,
+       MMC_BLK_RETRY,
+       MMC_BLK_ABORT,
+       MMC_BLK_DATA_ERR,
+       MMC_BLK_ECC_ERR,
+       MMC_BLK_NOMEDIUM,
+       MMC_BLK_NEW_REQUEST,
+};
+
 /* The number of MMC physical partitions.  These consist of:
  * boot partitions (2), general purpose partitions (4) in MMC v4.4.
  */
index 5bf7c2274fcb711138d517a131b2898d12f5b8a8..495d1336149c816518147a44a988cb88365aedff 100644 (file)
@@ -120,6 +120,7 @@ struct mmc_data {
        s32                     host_cookie;    /* host private data */
 };
 
+struct mmc_host;
 struct mmc_request {
        struct mmc_command      *sbc;           /* SET_BLOCK_COUNT for multiblock */
        struct mmc_command      *cmd;
@@ -128,9 +129,9 @@ struct mmc_request {
 
        struct completion       completion;
        void                    (*done)(struct mmc_request *);/* completion function */
+       struct mmc_host         *host;
 };
 
-struct mmc_host;
 struct mmc_card;
 struct mmc_async_req;
 
index c89a1bb87fa5b4a655a76da23a87602bd4738a92..523d570f58ad880bc478d17acdd6266f2b54e348 100644 (file)
@@ -170,6 +170,22 @@ struct mmc_slot {
        void *handler_priv;
 };
 
+/**
+ * mmc_context_info - synchronization details for mmc context
+ * @is_done_rcv                wake up reason was done request
+ * @is_new_req         wake up reason was new request
+ * @is_waiting_last_req        mmc context waiting for single running request
+ * @wait               wait queue
+ * @lock               lock to protect data fields
+ */
+struct mmc_context_info {
+       bool                    is_done_rcv;
+       bool                    is_new_req;
+       bool                    is_waiting_last_req;
+       wait_queue_head_t       wait;
+       spinlock_t              lock;
+};
+
 struct regulator;
 
 struct mmc_supply {
@@ -331,6 +347,7 @@ struct mmc_host {
        struct dentry           *debugfs_root;
 
        struct mmc_async_req    *areq;          /* active async req */
+       struct mmc_context_info context_info;   /* async synchronization info */
 
 #ifdef CONFIG_FAIL_MMC_REQUEST
        struct fault_attr       fail_mmc_request;