Skip to contentSkip to conversation

fix random SD/MMC card recognition failures on ARM Versatile

PATCH In review Vitaly Wool 2 participants 6 messages Nov 23, 2006 – Nov 24, 2006 MMC, SD and SDIO

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fix random SD/MMC card recognition failures on ARM Versatile

Hello Pierre,

currently sometimes the SD/MMC card inserted results in recognition failure on ARM Versatile board:

<<<Plug in MMC card>>>

root@versatile:~# mmcblk0: mmc0:0001 SDMB-32 31360KiB
 mmcblk0:<3>mmcblk0: error 3 transferring data
end_request: I/O error, dev mmcblk0, sector 0
Buffer I/O error on device mmcblk0, logical block 0
mmcblk0: error 3 transferring data
end_request: I/O error, dev mmcblk0, sector 0
Buffer I/O error on device mmcblk0, logical block 0
 unable to read partition table

This patch fixes the problem.

 drivers/mmc/mmci.c |    2 ++
 1 file changed, 2 insertions(+)
drivers/mmc/mmci.c +2 -0
Index: linux-2.6.18/drivers/mmc/mmci.c
===================================================================
@@ -41,6 +41,8 @@ static void
 mmci_request_end(struct mmci_host *host, struct mmc_request *mrq)
 {
 	writel(0, host->base + MMCICOMMAND);
+	writel(0, host->base + MMCIDATACTRL);
+	writel(0, host->base + MMCIMASK1);
 
 	host->mrq = NULL;
 	host->cmd = NULL;
#1 lore raw
Quoted 15 lines · Vitaly Wool

Hello Pierre,

currently sometimes the SD/MMC card inserted results in recognition failure on ARM Versatile board:

   <<<Plug in MMC card>>>
root@versatile:~# mmcblk0: mmc0:0001 SDMB-32 31360KiB
 mmcblk0:<3>mmcblk0: error 3 transferring data
end_request: I/O error, dev mmcblk0, sector 0
Buffer I/O error on device mmcblk0, logical block 0
mmcblk0: error 3 transferring data
end_request: I/O error, dev mmcblk0, sector 0
Buffer I/O error on device mmcblk0, logical block 0
 unable to read partition table
Quoted 1 line

This patch fixes the problem.

Doubtful. mmci_stop_data() already does this, which will be called immediately prior to mmci_request_end(). So you're doubling up the writes to registers again.

Since this is not the first occurance that you've had to do this with your board (the other being the SIC) I suggest that your board is faulty in some way, causing writes to registers to be occasionally dropped.

Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:
#2 lore raw
Quoted 3 lines · Russell King

Doubtful. mmci_stop_data() already does this, which will be called immediately prior to mmci_request_end(). So you're doubling up the writes to registers again.

There's the case (mmci_cmd_irq) where mmc_stop_data is not called prior to mmci_request_end(), so it's not that simple.

Quoted 4 lines

Since this is not the first occurance that you've had to do this with your board (the other being the SIC) I suggest that your board is faulty in some way, causing writes to registers to be occasionally dropped.

I can't 100% guarantee that it's not the case, but the problem has been reproduced by at least 2 people on 2 different boards and on 2 different sides of the Atlantic. So I'd suspect there's either a SW problem or a HW revision problem, at least.

Vitaly

#3 lore raw
Quoted 7 lines · Vitaly Wool
Quoted 3 lines · Russell King

Doubtful. mmci_stop_data() already does this, which will be called immediately prior to mmci_request_end(). So you're doubling up the writes to registers again.

There's the case (mmci_cmd_irq) where mmc_stop_data is not called prior to mmci_request_end(), so it's not that simple.

Ah, I see it. In that case we need to call mmc_stop_data() when we're ending the initial command due to an error. IOW, like this:

drivers/mmc/mmci.c +4 -0
index 828503c..5ad0259 100644
@@ -42,6 +42,8 @@ mmci_request_end(struct mmci_host *host,
 {
 	writel(0, host->base + MMCICOMMAND);
 
+	BUG_ON(host->data);
+
 	host->mrq = NULL;
 	host->cmd = NULL;
 
@@ -198,6 +200,8 @@ mmci_cmd_irq(struct mmci_host *host, str
 	}
 
 	if (!cmd->data || cmd->error != MMC_ERR_NONE) {
+		if (host->data)
+			mmci_stop_data(host);
 		mmci_request_end(host, cmd->mrq);
 	} else if (!(cmd->data->flags & MMC_DATA_READ)) {
 		mmci_start_data(host, cmd->data);
Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:
#4 lore raw
Quoted 11 lines · Russell King
Quoted 7 lines · Vitaly Wool
Quoted 3 lines · Russell King

Doubtful. mmci_stop_data() already does this, which will be called immediately prior to mmci_request_end(). So you're doubling up the writes to registers again.

There's the case (mmci_cmd_irq) where mmc_stop_data is not called prior to mmci_request_end(), so it's not that simple.

Ah, I see it. In that case we need to call mmc_stop_data() when we're ending the initial command due to an error. IOW, like this:

I'll also add that with the way we handle the MMCI, it is highly likely that you _will_ see FIFO errors from time to time on this platform.

The problem is that we don't have DMA up and running on this platform, so we are entirely at the mercy of interrupt-driven PIO. In addition, the MMCI FIFOs must be read _before_ they completely fill to avoid overrun errors. Coupling these two facts together, it's easy to see that interrupt latency is _critical_ to avoiding FIFO overruns (error 3).

In general, if you do _anything_ with the board while it's trying to access MMC cards, you will probably get some FIFO overruns.

There are three solutions:

1. Lower the maximum clock rate that the MMCI will allow, eg:
insmod mmci fmax=257816

2. Avoid all other system activity while MMC is being accessed.

3. Someone needs to _sanely_ implement DMA on this platform.

Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:
#5 lore raw
In reply to Russell King
Quoted 2 lines · On Thu, 23 Nov 2006 19:42:36 +0000 Russell King <[email protected]>…

Ah, I see it. In that case we need to call mmc_stop_data() when we're ending the initial command due to an error. IOW, like this:

<snip>

I'd suggest arranging that in a bit different way. It looks like it works better when MMCIDATACTRL/MMCIMASK1 are cleared after MMCICOMMAND (and I think it makes more sense to clear the command register first, thus we have less change to get spurious interrupts after MMCIMASK1 is set).

drivers/mmc/mmci.c +10 -9
index 828503c..afbb63b 100644
@@ -37,11 +37,21 @@ #define DBG(host,fmt,args...)	\
 
 static unsigned int fmax = 515633;
 
+static void mmci_stop_data(struct mmci_host *host)
+{
+	writel(0, host->base + MMCIDATACTRL);
+	writel(0, host->base + MMCIMASK1);
+	host->data = NULL;
+}
+
 static void
 mmci_request_end(struct mmci_host *host, struct mmc_request *mrq)
 {
 	writel(0, host->base + MMCICOMMAND);
 
+	if (host->data)
+		mmci_stop_data(host);
+
 	host->mrq = NULL;
 	host->cmd = NULL;
 
@@ -57,13 +67,6 @@ mmci_request_end(struct mmci_host *host,
 	spin_lock(&host->lock);
 }
 
-static void mmci_stop_data(struct mmci_host *host)
-{
-	writel(0, host->base + MMCIDATACTRL);
-	writel(0, host->base + MMCIMASK1);
-	host->data = NULL;
-}
-
 static void mmci_start_data(struct mmci_host *host, struct mmc_data *data)
 {
 	unsigned int datactrl, timeout, irqmask;
@@ -168,8 +171,6 @@ mmci_data_irq(struct mmci_host *host, st
 			flush_dcache_page(host->sg_ptr->page);
 	}
 	if (status & MCI_DATAEND) {
-		mmci_stop_data(host);
-
 		if (!data->stop) {
 			mmci_request_end(host, data->mrq);
 		} else {

Vitaly