MMC: PL180: Fix infinite loop with VExpress extended fifo implementation

One possible fix is to add some build time configuration to change SDI_FIFO_BURST_SIZE for the new implementation.

Hi Tixy,

Quoted 2 lines

One possible fix is to add some build time configuration to change SDI_FIFO_BURST_SIZE for the new implementation.

You can also detect the configuration in runtime, basing on PeriphID:

http://infocenter.arm.com/help/topic/com.arm.doc.ddi0172a/i1024149.html

Configuration == 0 (ID == 0x00041180) -> FIFO length = 16 * 4 Configuration == 1 (ID == 0x01041180) -> FIFO length = 128 * 4

That's useful to know. The PL180 code is also used for U8500, I don't know if that implements the peripheral ID register; though I guess any probing could be limited to vexpress anyway.

Quoted 3 lines

That's useful to know. The PL180 code is also used for U8500, I don't know if that implements the peripheral ID register; though I guess any probing could be limited to vexpress anyway.

STE have the same "problems" with FIFO size, see drivers/mmc/host/mmci.c in kernel sources:

From: Jon Medhurst <[email protected]>

The new IO FPGA implementation for Versatile Express contains an MMCI (PL180) cell with the FIFO extended to 128 words. This causes the read_bytes() function to go into an infinite loop; as it will wait for for the half-full signal (SDI_STA_RXFIFOBR) if there are more than 8 words remaining (SDI_FIFO_BURST_SIZE), but it won't receive this signal once there are fewer than 64 words left to transfer.

One possible fix is to add some build time configuration to change SDI_FIFO_BURST_SIZE for the new implementation. However, the problematic code only seems to exist as a small performance optimisation, so the solution implemented by this patch is to simply remove it. The error checking following the loop is also removed as this will be handled by code further down the function.

Cc Andy Fleming <[email protected]>

Signed-off-by Jon Medhurst <[email protected]>

---

If it is desirable to keep the burst read optimisation, then an alternative solution would be to keep the loop and add an if clause to do a single read if the fifo doesn't have enough for a burst.

---
 drivers/mmc/arm_pl180_mmci.c |   26 --------------------------
 1 files changed, 0 insertions(+), 26 deletions(-)

drivers/mmc/arm_pl180_mmci.c +0 -26

index ed296ee..e6467a2 100644
@@ -111,7 +111,6 @@ static int do_command(struct mmc *dev, struct mmc_cmd *cmd)
 static int read_bytes(struct mmc *dev, u32 *dest, u32 blkcount, u32 blksize)
 {
        u32 *tempbuff = dest;
-       int i;
        u64 xfercount = blkcount * blksize;
        struct mmc_host *host = dev->priv;
        u32 status, status_err;

@@ -121,31 +120,6 @@ static int read_bytes(struct mmc *dev, u32 *dest, u32 blkcount, u32 blksize)
        status = readl(&host->base->status);
        status_err = status & (SDI_STA_DCRCFAIL | SDI_STA_DTIMEOUT |
                               SDI_STA_RXOVERR);
-       while (!status_err &&
-              (xfercount >= SDI_FIFO_BURST_SIZE * sizeof(u32))) {
-               if (status & SDI_STA_RXFIFOBR) {
-                       for (i = 0; i < SDI_FIFO_BURST_SIZE; i++)
-                               *(tempbuff + i) = readl(&host->base->fifo);
-                       tempbuff += SDI_FIFO_BURST_SIZE;
-                       xfercount -= SDI_FIFO_BURST_SIZE * sizeof(u32);
-               }
-               status = readl(&host->base->status);
-               status_err = status &
-                       (SDI_STA_DCRCFAIL | SDI_STA_DTIMEOUT | SDI_STA_RXOVERR);
-       }
-
-       if (status & SDI_STA_DTIMEOUT) {
-               printf("Read data timed out, xfercount: %llu, status: 0x%08X\n",
-                       xfercount, status);
-               return -ETIMEDOUT;
-       } else if (status & SDI_STA_DCRCFAIL) {
-               printf("Read data blk CRC error: 0x%x\n", status);
-               return -EILSEQ;
-       } else if (status & SDI_STA_RXOVERR) {
-               printf("Read data RX overflow error\n");
-               return -EIO;
-       }
-
        while ((!status_err) && (xfercount >= sizeof(u32))) {
                if (status & SDI_STA_RXDAVL) {
                        *(tempbuff) = readl(&host->base->fifo);

Cc Andy Fleming <[email protected]>

Signed-off-by Jon Medhurst <[email protected]>

?drivers/mmc/arm_pl180_mmci.c | ? 26 -------------------------- ?1 files changed, 0 insertions(+), 26 deletions(-)

drivers/mmc/arm_pl180_mmci.c +0 -0

index ed296ee..e6467a2 100644

The new IO FPGA implementation for Versatile Express contains an MMCI (PL180) cell with the FIFO extended to 128 words. This causes the read_bytes() function to go into an infinite loop; as it will wait for for the half-full signal (SDI_STA_RXFIFOBR) if there are more than 8 words remaining (SDI_FIFO_BURST_SIZE), but it won't receive this signal once there are fewer than 64 words left to transfer.

One possible fix is to add some build time configuration to change SDI_FIFO_BURST_SIZE for the new implementation. However, the problematic code only seems to exist as a small performance optimisation, so the solution implemented by this patch is to simply remove it. The error checking following the loop is also removed as this will be handled by code further down the function.

Cc Andy Fleming <[email protected]>

Signed-off-by Jon Medhurst <[email protected]>