]> err.no Git - linux-2.6/commitdiff
mmc: bounce requests for simple hosts
authorPierre Ossman <drzeus@drzeus.cx>
Fri, 11 May 2007 22:26:16 +0000 (00:26 +0200)
committerPierre Ossman <drzeus@drzeus.cx>
Mon, 9 Jul 2007 19:22:53 +0000 (21:22 +0200)
Some hosts cannot do scatter/gather in hardware. Since not doing sg
is such a big performance hit, we (optionally) bounce the requests
to a simple linear buffer that we hand over to the driver.

Signed-off-by: Pierre Ossman <drzeus@drzeus.cx>
drivers/mmc/card/Kconfig
drivers/mmc/card/block.c
drivers/mmc/card/queue.c
drivers/mmc/card/queue.h

index 9320a8c73239ceff909e30a045bb06aec60bd5ee..a49cb9737cd858c8f36679249b29f86b565426f2 100644 (file)
@@ -14,3 +14,21 @@ config MMC_BLOCK
          mount the filesystem. Almost everyone wishing MMC support
          should say Y or M here.
 
+config MMC_BLOCK_BOUNCE
+       bool "Use bounce buffer for simple hosts"
+       depends on MMC_BLOCK
+       default y
+       help
+         SD/MMC is a high latency protocol where it is crucial to
+         send large requests in order to get high performance. Many
+         controllers, however, are restricted to continuous memory
+         (i.e. they can't do scatter-gather), something the kernel
+         rarely can provide.
+
+         Say Y here to help these restricted hosts by bouncing
+         requests back and forth from a large buffer. You will get
+         a big performance gain at the cost of up to 64 KiB of
+         physical memory.
+
+         If unsure, say Y here.
+
index 540ff4bea54c6f8eb3366b6066393dec6567ae8e..cbd4b6e3e17c6a30cb91fc24aff63fe12dd14384 100644 (file)
@@ -262,7 +262,9 @@ static int mmc_blk_issue_rq(struct mmc_queue *mq, struct request *req)
                }
 
                brq.data.sg = mq->sg;
-               brq.data.sg_len = blk_rq_map_sg(req->q, req, brq.data.sg);
+               brq.data.sg_len = mmc_queue_map_sg(mq);
+
+               mmc_queue_bounce_pre(mq);
 
                if (brq.data.blocks !=
                    (req->nr_sectors >> (md->block_bits - 9))) {
@@ -279,6 +281,9 @@ static int mmc_blk_issue_rq(struct mmc_queue *mq, struct request *req)
                }
 
                mmc_wait_for_req(card->host, &brq.mrq);
+
+               mmc_queue_bounce_post(mq);
+
                if (brq.cmd.error) {
                        printk(KERN_ERR "%s: error %d sending read/write command\n",
                               req->rq_disk->disk_name, brq.cmd.error);
index dd97bc798409eb0bdec4d5a413c217c173a642e4..4fb2089dc6900cb1399496defcf8184b106cdbd0 100644 (file)
@@ -17,6 +17,8 @@
 #include <linux/mmc/host.h>
 #include "queue.h"
 
+#define MMC_QUEUE_BOUNCESZ     65536
+
 #define MMC_QUEUE_SUSPENDED    (1 << 0)
 
 /*
@@ -118,6 +120,7 @@ int mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card, spinlock_t *lock
        struct mmc_host *host = card->host;
        u64 limit = BLK_BOUNCE_HIGH;
        int ret;
+       unsigned int bouncesz;
 
        if (mmc_dev(host)->dma_mask && *mmc_dev(host)->dma_mask)
                limit = *mmc_dev(host)->dma_mask;
@@ -127,21 +130,61 @@ int mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card, spinlock_t *lock
        if (!mq->queue)
                return -ENOMEM;
 
-       blk_queue_prep_rq(mq->queue, mmc_prep_request);
-       blk_queue_bounce_limit(mq->queue, limit);
-       blk_queue_max_sectors(mq->queue, host->max_req_size / 512);
-       blk_queue_max_phys_segments(mq->queue, host->max_phys_segs);
-       blk_queue_max_hw_segments(mq->queue, host->max_hw_segs);
-       blk_queue_max_segment_size(mq->queue, host->max_seg_size);
-
        mq->queue->queuedata = mq;
        mq->req = NULL;
 
-       mq->sg = kmalloc(sizeof(struct scatterlist) * host->max_phys_segs,
-                        GFP_KERNEL);
-       if (!mq->sg) {
-               ret = -ENOMEM;
-               goto cleanup_queue;
+       blk_queue_prep_rq(mq->queue, mmc_prep_request);
+
+#ifdef CONFIG_MMC_BLOCK_BOUNCE
+       if (host->max_hw_segs == 1) {
+               bouncesz = MMC_QUEUE_BOUNCESZ;
+
+               if (bouncesz > host->max_req_size)
+                       bouncesz = host->max_req_size;
+               if (bouncesz > host->max_seg_size)
+                       bouncesz = host->max_seg_size;
+
+               mq->bounce_buf = kmalloc(bouncesz, GFP_KERNEL);
+               if (!mq->bounce_buf) {
+                       printk(KERN_WARNING "%s: unable to allocate "
+                               "bounce buffer\n", mmc_card_name(card));
+               } else {
+                       blk_queue_bounce_limit(mq->queue, BLK_BOUNCE_HIGH);
+                       blk_queue_max_sectors(mq->queue, bouncesz / 512);
+                       blk_queue_max_phys_segments(mq->queue, bouncesz / 512);
+                       blk_queue_max_hw_segments(mq->queue, bouncesz / 512);
+                       blk_queue_max_segment_size(mq->queue, bouncesz);
+
+                       mq->sg = kmalloc(sizeof(struct scatterlist),
+                               GFP_KERNEL);
+                       if (!mq->sg) {
+                               ret = -ENOMEM;
+                               goto free_bounce_buf;
+                       }
+
+                       mq->bounce_sg = kmalloc(sizeof(struct scatterlist) *
+                               bouncesz / 512, GFP_KERNEL);
+                       if (!mq->bounce_sg) {
+                               ret = -ENOMEM;
+                               goto free_sg;
+                       }
+               }
+       }
+#endif
+
+       if (!mq->bounce_buf) {
+               blk_queue_bounce_limit(mq->queue, limit);
+               blk_queue_max_sectors(mq->queue, host->max_req_size / 512);
+               blk_queue_max_phys_segments(mq->queue, host->max_phys_segs);
+               blk_queue_max_hw_segments(mq->queue, host->max_hw_segs);
+               blk_queue_max_segment_size(mq->queue, host->max_seg_size);
+
+               mq->sg = kmalloc(sizeof(struct scatterlist) *
+                       host->max_phys_segs, GFP_KERNEL);
+               if (!mq->sg) {
+                       ret = -ENOMEM;
+                       goto cleanup_queue;
+               }
        }
 
        init_MUTEX(&mq->thread_sem);
@@ -149,14 +192,21 @@ int mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card, spinlock_t *lock
        mq->thread = kthread_run(mmc_queue_thread, mq, "mmcqd");
        if (IS_ERR(mq->thread)) {
                ret = PTR_ERR(mq->thread);
-               goto free_sg;
+               goto free_bounce_sg;
        }
 
        return 0;
-
+ free_bounce_sg:
+       if (mq->bounce_sg)
+               kfree(mq->bounce_sg);
+       mq->bounce_sg = NULL;
  free_sg:
        kfree(mq->sg);
        mq->sg = NULL;
+ free_bounce_buf:
+       if (mq->bounce_buf)
+               kfree(mq->bounce_buf);
+       mq->bounce_buf = NULL;
  cleanup_queue:
        blk_cleanup_queue(mq->queue);
        return ret;
@@ -178,9 +228,17 @@ void mmc_cleanup_queue(struct mmc_queue *mq)
        /* Then terminate our worker thread */
        kthread_stop(mq->thread);
 
+       if (mq->bounce_sg)
+               kfree(mq->bounce_sg);
+       mq->bounce_sg = NULL;
+
        kfree(mq->sg);
        mq->sg = NULL;
 
+       if (mq->bounce_buf)
+               kfree(mq->bounce_buf);
+       mq->bounce_buf = NULL;
+
        blk_cleanup_queue(mq->queue);
 
        mq->card = NULL;
@@ -231,3 +289,108 @@ void mmc_queue_resume(struct mmc_queue *mq)
        }
 }
 
+static void copy_sg(struct scatterlist *dst, unsigned int dst_len,
+       struct scatterlist *src, unsigned int src_len)
+{
+       unsigned int chunk;
+       char *dst_buf, *src_buf;
+       unsigned int dst_size, src_size;
+
+       dst_buf = NULL;
+       src_buf = NULL;
+       dst_size = 0;
+       src_size = 0;
+
+       while (src_len) {
+               BUG_ON(dst_len == 0);
+
+               if (dst_size == 0) {
+                       dst_buf = page_address(dst->page) + dst->offset;
+                       dst_size = dst->length;
+               }
+
+               if (src_size == 0) {
+                       src_buf = page_address(src->page) + src->offset;
+                       src_size = src->length;
+               }
+
+               chunk = min(dst_size, src_size);
+
+               memcpy(dst_buf, src_buf, chunk);
+
+               dst_buf += chunk;
+               src_buf += chunk;
+               dst_size -= chunk;
+               src_size -= chunk;
+
+               if (dst_size == 0) {
+                       dst++;
+                       dst_len--;
+               }
+
+               if (src_size == 0) {
+                       src++;
+                       src_len--;
+               }
+       }
+}
+
+unsigned int mmc_queue_map_sg(struct mmc_queue *mq)
+{
+       unsigned int sg_len;
+
+       if (!mq->bounce_buf)
+               return blk_rq_map_sg(mq->queue, mq->req, mq->sg);
+
+       BUG_ON(!mq->bounce_sg);
+
+       sg_len = blk_rq_map_sg(mq->queue, mq->req, mq->bounce_sg);
+
+       mq->bounce_sg_len = sg_len;
+
+       /*
+        * Shortcut in the event we only get a single entry.
+        */
+       if (sg_len == 1) {
+               memcpy(mq->sg, mq->bounce_sg, sizeof(struct scatterlist));
+               return 1;
+       }
+
+       mq->sg[0].page = virt_to_page(mq->bounce_buf);
+       mq->sg[0].offset = offset_in_page(mq->bounce_buf);
+       mq->sg[0].length = 0;
+
+       while (sg_len) {
+               mq->sg[0].length += mq->bounce_sg[sg_len - 1].length;
+               sg_len--;
+       }
+
+       return 1;
+}
+
+void mmc_queue_bounce_pre(struct mmc_queue *mq)
+{
+       if (!mq->bounce_buf)
+               return;
+
+       if (mq->bounce_sg_len == 1)
+               return;
+       if (rq_data_dir(mq->req) != WRITE)
+               return;
+
+       copy_sg(mq->sg, 1, mq->bounce_sg, mq->bounce_sg_len);
+}
+
+void mmc_queue_bounce_post(struct mmc_queue *mq)
+{
+       if (!mq->bounce_buf)
+               return;
+
+       if (mq->bounce_sg_len == 1)
+               return;
+       if (rq_data_dir(mq->req) != READ)
+               return;
+
+       copy_sg(mq->bounce_sg, mq->bounce_sg_len, mq->sg, 1);
+}
+
index 1590b3f3f1f7d8d54fbdf1eb1a27c6f25893205e..64e66e0d4994a9176662354cd123fb0cee2315a0 100644 (file)
@@ -14,6 +14,9 @@ struct mmc_queue {
        void                    *data;
        struct request_queue    *queue;
        struct scatterlist      *sg;
+       char                    *bounce_buf;
+       struct scatterlist      *bounce_sg;
+       unsigned int            bounce_sg_len;
 };
 
 extern int mmc_init_queue(struct mmc_queue *, struct mmc_card *, spinlock_t *);
@@ -21,4 +24,8 @@ extern void mmc_cleanup_queue(struct mmc_queue *);
 extern void mmc_queue_suspend(struct mmc_queue *);
 extern void mmc_queue_resume(struct mmc_queue *);
 
+extern unsigned int mmc_queue_map_sg(struct mmc_queue *);
+extern void mmc_queue_bounce_pre(struct mmc_queue *);
+extern void mmc_queue_bounce_post(struct mmc_queue *);
+
 #endif