Revert "mmc: alcor: enable DMA transfer of large buffers"
authorDaniel Drake <drake@endlessm.com>
Mon, 29 Apr 2019 05:14:25 +0000 (13:14 +0800)
committerUlf Hansson <ulf.hansson@linaro.org>
Mon, 6 May 2019 09:55:39 +0000 (11:55 +0200)
This reverts commit 57ebb96c293da9f0ec56aba13c5541269a5c10b1.

Usage of the DMA page iterator was problematic here because
we were not considering offset & length of entries in the scatterlist.

Also, after further discussion, the suggested revised approach is much
more similar to the driver implementation before this commit was
applied, so revert it.

Signed-off-by: Daniel Drake <drake@endlessm.com>
Signed-off-by: Ulf Hansson <ulf.hansson@linaro.org>
drivers/mmc/host/alcor.c

index ddcdab14707c8bfbc0e8a2c9cf9d1f4d06c9ef83..3e2efbf3cf8cdef727661b782ef2dda1a3fe4839 100644 (file)
@@ -54,9 +54,9 @@ struct alcor_sdmmc_host {
        struct delayed_work timeout_work;
 
        struct sg_mapping_iter sg_miter;        /* SG state for PIO */
-       struct sg_dma_page_iter sg_diter;       /* SG state for DMA */
        struct scatterlist *sg;
        unsigned int blocks;            /* remaining PIO blocks */
+       int sg_count;
 
        u32                     irq_status_sd;
        unsigned char           cur_power_mode;
@@ -117,19 +117,30 @@ static void alcor_reset(struct alcor_sdmmc_host *host, u8 val)
        dev_err(host->dev, "%s: timeout\n", __func__);
 }
 
-/*
- * Perform DMA I/O of a single page.
- */
 static void alcor_data_set_dma(struct alcor_sdmmc_host *host)
 {
        struct alcor_pci_priv *priv = host->alcor_pci;
-       dma_addr_t addr;
+       u32 addr;
+
+       if (!host->sg_count)
+               return;
 
-       if (!__sg_page_iter_dma_next(&host->sg_diter))
+       if (!host->sg) {
+               dev_err(host->dev, "have blocks, but no SG\n");
                return;
+       }
 
-       addr = sg_page_iter_dma_address(&host->sg_diter);
-       alcor_write32(priv, (u32) addr, AU6601_REG_SDMA_ADDR);
+       if (!sg_dma_len(host->sg)) {
+               dev_err(host->dev, "DMA SG len == 0\n");
+               return;
+       }
+
+
+       addr = (u32)sg_dma_address(host->sg);
+
+       alcor_write32(priv, addr, AU6601_REG_SDMA_ADDR);
+       host->sg = sg_next(host->sg);
+       host->sg_count--;
 }
 
 static void alcor_trigger_data_transfer(struct alcor_sdmmc_host *host)
@@ -142,29 +153,12 @@ static void alcor_trigger_data_transfer(struct alcor_sdmmc_host *host)
                ctrl |= AU6601_DATA_WRITE;
 
        if (data->host_cookie == COOKIE_MAPPED) {
-               /*
-                * For DMA transfers, this function is called just once,
-                * at the start of the operation. The hardware can only
-                * perform DMA I/O on a single page at a time, so here
-                * we kick off the transfer with the first page, and expect
-                * subsequent pages to be transferred upon IRQ events
-                * indicating that the single-page DMA was completed.
-                */
-               __sg_page_iter_start(&host->sg_diter.base, data->sg,
-                                    data->sg_len, 0);
-
                alcor_data_set_dma(host);
                ctrl |= AU6601_DATA_DMA_MODE;
                host->dma_on = 1;
-               alcor_write32(priv, data->blksz * data->blocks,
+               alcor_write32(priv, data->sg_count * 0x1000,
                               AU6601_REG_BLOCK_SIZE);
        } else {
-               /*
-                * For PIO transfers, we break down each operation
-                * into several sector-sized transfers. When one sector has
-                * complete, the IRQ handler will call this function again
-                * to kick off the transfer of the next sector.
-                */
                alcor_write32(priv, data->blksz, AU6601_REG_BLOCK_SIZE);
        }
 
@@ -239,8 +233,9 @@ static void alcor_prepare_data(struct alcor_sdmmc_host *host,
        host->data->bytes_xfered = 0;
        host->blocks = data->blocks;
        host->sg = data->sg;
+       host->sg_count = data->sg_count;
        dev_dbg(host->dev, "prepare DATA: sg %i, blocks: %i\n",
-                       data->sg_count, host->blocks);
+                       host->sg_count, host->blocks);
 
        if (data->host_cookie != COOKIE_MAPPED)
                alcor_prepare_sg_miter(host);
@@ -489,6 +484,9 @@ static int alcor_data_irq_done(struct alcor_sdmmc_host *host, u32 intmask)
                alcor_trf_block_pio(host, false);
                return 1;
        case AU6601_INT_DMA_END:
+               if (!host->sg_count)
+                       break;
+
                alcor_data_set_dma(host);
                break;
        default:
@@ -525,7 +523,8 @@ static void alcor_data_irq_thread(struct alcor_sdmmc_host *host, u32 intmask)
        if (alcor_data_irq_done(host, intmask))
                return;
 
-       if ((intmask & AU6601_INT_DATA_END) || !host->blocks || host->dma_on)
+       if ((intmask & AU6601_INT_DATA_END) || !host->blocks ||
+           (host->dma_on && !host->sg_count))
                alcor_finish_data(host);
 }
 
@@ -763,7 +762,8 @@ static void alcor_pre_req(struct mmc_host *mmc,
        struct alcor_sdmmc_host *host = mmc_priv(mmc);
        struct mmc_data *data = mrq->data;
        struct mmc_command *cmd = mrq->cmd;
-       unsigned int sg_len;
+       struct scatterlist *sg;
+       unsigned int i, sg_len;
 
        if (!data || !cmd)
                return;
@@ -785,6 +785,11 @@ static void alcor_pre_req(struct mmc_host *mmc,
        if (data->blksz & 3)
                return;
 
+       for_each_sg(data->sg, sg, data->sg_len, i) {
+               if (sg->length != AU6601_MAX_DMA_BLOCK_SIZE)
+                       return;
+       }
+
        /* This data might be unmapped at this time */
 
        sg_len = dma_map_sg(host->dev, data->sg, data->sg_len,
@@ -1032,13 +1037,26 @@ static void alcor_init_mmc(struct alcor_sdmmc_host *host)
        mmc->caps2 = MMC_CAP2_NO_SDIO;
        mmc->ops = &alcor_sdc_ops;
 
-       /*
-        * Enable large requests through iteration of scatterlist pages.
-        * Limit to 240 sectors per request like the original vendor driver.
+       /* The hardware does DMA data transfer of 4096 bytes to/from a single
+        * buffer address. Scatterlists are not supported, but upon DMA
+        * completion (signalled via IRQ), the original vendor driver does
+        * then immediately set up another DMA transfer of the next 4096
+        * bytes.
+        *
+        * This means that we need to handle the I/O in 4096 byte chunks.
+        * Lacking a way to limit the sglist entries to 4096 bytes, we instead
+        * impose that only one segment is provided, with maximum size 4096,
+        * which also happens to be the minimum size. This means that the
+        * single-entry sglist handled by this driver can be handed directly
+        * to the hardware, nice and simple.
+        *
+        * Unfortunately though, that means we only do 4096 bytes I/O per
+        * MMC command. A future improvement would be to make the driver
+        * accept sg lists and entries of any size, and simply iterate
+        * through them 4096 bytes at a time.
         */
-       mmc->max_segs = 64;
-       mmc->max_seg_size = 240 * 512;
-       mmc->max_blk_count = 240;
+       mmc->max_segs = AU6601_MAX_DMA_SEGMENTS;
+       mmc->max_seg_size = AU6601_MAX_DMA_BLOCK_SIZE;
        mmc->max_req_size = mmc->max_seg_size;
 }