Re: [PATCH/RFC 1/2] 5200: improve i2c bus error recovery
From: Ira W. Snyder <hidden>
Date: 2010-02-16 19:49:22
Also in:
linuxppc-dev
On Fri, Jan 22, 2010 at 09:17:55PM +0100, Albrecht Dreß wrote:
Improve the recovery of the MPC5200B's I2C bus from errors like bus hangs. Signed-off-by: Albrecht Dreß <redacted> --- This patch introduces several improvements to the MPC5200B's I2C driver as to improve the recovery from error conditions I encountered when testing a custom board with several I2C devices attached (eeprom, io expander, rtc, sensors). The error conditions included cases where the bus if logic of one slave apparently went south, blocking the bus completely. My fixes include: 1. make the bus timeout configurable in fsl_i2c_probe(); the default of one second is *way* too long for my use case; 2. if a timeout condition occurs in mpc_xfer(), mpc_i2c_fixup() the bus if *any* of the CF, BB and RXAK flags in the MSR is 1. I actually saw different combinations with hangs, not only all three set;
Hello Albrecht, I see this exact hang on a MPC8349EA board. I poll my i2c sensors every 500ms, and it takes around 12 hours to produce a hang. The usual hang has (CF | BB) set, however I have seen a hang with just BB (only once so far in about 2 weeks). I think the fixup should be run on 8349 as well, if not all processors. I'm happy to test patches. I have a way to reliably trigger a lockup, using another master on the i2c bus.
quoted hunk ↗ jump to hunk
3. improve the fixup procedure by calculating the timing needed from the real (configured) bus clock, calculated in mpc_i2c_setclock_52xx(). Furthermore, I issue 9 instead of one cycle, as I experienced cases where the single one is not enough (found this tip in a forum). As a side effect, the new scheme needs only 81us @375kHz bus clock instead of 150us. I recorded waveforms for 18.4kHz, 85.9kHz and 375kHz, all looking fine, which I can provide if anyone is interested. Open questions: - is the approach correct at all, in particular the interpretation of the flags (#2)? - could this code also be used on non-5200 processors?--- linux-2.6.32-orig/drivers/i2c/busses/i2c-mpc.c 2009-12-03 04:51:21.000000000 +0100 +++ linux-2.6.32/drivers/i2c/busses/i2c-mpc.c 2010-01-22 16:05:13.000000000 +0100@@ -59,6 +59,7 @@ struct mpc_i2c { wait_queue_head_t queue; struct i2c_adapter adap; int irq; + u32 real_clk; }; struct mpc_i2c_divider {@@ -97,16 +98,32 @@ static irqreturn_t mpc_i2c_isr(int irq, */ static void mpc_i2c_fixup(struct mpc_i2c *i2c) { - writeccr(i2c, 0); - udelay(30); - writeccr(i2c, CCR_MEN); - udelay(30); - writeccr(i2c, CCR_MSTA | CCR_MTX); - udelay(30); - writeccr(i2c, CCR_MSTA | CCR_MTX | CCR_MEN); - udelay(30); - writeccr(i2c, CCR_MEN); - udelay(30); + if (i2c->real_clk == 0) { + writeccr(i2c, 0); + udelay(30); + writeccr(i2c, CCR_MEN); + udelay(30); + writeccr(i2c, CCR_MSTA | CCR_MTX); + udelay(30); + writeccr(i2c, CCR_MSTA | CCR_MTX | CCR_MEN); + udelay(30); + writeccr(i2c, CCR_MEN); + udelay(30); + } else { + int k; + u32 delay_val = 1000000 / i2c->real_clk + 1; + + if (delay_val < 2) + delay_val = 2; + + for (k = 9; k; k--) { + writeccr(i2c, 0); + writeccr(i2c, CCR_MSTA | CCR_MTX | CCR_MEN); + udelay(delay_val); + writeccr(i2c, CCR_MEN); + udelay(delay_val << 1); + } + } }
The old sequence has always un-hung the bus for me. Yours might be better. I'll try it next time the bus wedges up on me.
quoted hunk ↗ jump to hunk
static int i2c_wait(struct mpc_i2c *i2c, unsigned timeout, int writing)@@ -186,15 +203,18 @@ static const struct mpc_i2c_divider mpc_ {10240, 0x9d}, {12288, 0x9e}, {15360, 0x9f} }; -int mpc_i2c_get_fdr_52xx(struct device_node *node, u32 clock, int prescaler) +int mpc_i2c_get_fdr_52xx(struct device_node *node, u32 clock, int prescaler, + u32 *real_clk) { const struct mpc_i2c_divider *div = NULL; unsigned int pvr = mfspr(SPRN_PVR); u32 divider; int i; - if (!clock) + if (!clock) { + *real_clk = 0; return -EINVAL; + } /* Determine divider value */ divider = mpc5xxx_get_bus_frequency(node) / clock;@@ -212,7 +232,8 @@ int mpc_i2c_get_fdr_52xx(struct device_n break; } - return div ? (int)div->fdr : -EINVAL; + *real_clk = mpc5xxx_get_bus_frequency(node) / div->divider; + return (int)div->fdr; } static void mpc_i2c_setclock_52xx(struct device_node *node,@@ -221,13 +242,14 @@ static void mpc_i2c_setclock_52xx(struct { int ret, fdr; - ret = mpc_i2c_get_fdr_52xx(node, clock, prescaler); + ret = mpc_i2c_get_fdr_52xx(node, clock, prescaler, &i2c->real_clk); fdr = (ret >= 0) ? ret : 0x3f; /* backward compatibility */ writeb(fdr & 0xff, i2c->base + MPC_I2C_FDR); if (ret >= 0) - dev_info(i2c->dev, "clock %d Hz (fdr=%d)\n", clock, fdr); + dev_info(i2c->dev, "clock %u Hz (fdr=%d)\n", i2c->real_clk, + fdr); } #else /* !CONFIG_PPC_MPC52xx */ static void mpc_i2c_setclock_52xx(struct device_node *node,@@ -446,10 +468,14 @@ static int mpc_xfer(struct i2c_adapter * return -EINTR; } if (time_after(jiffies, orig_jiffies + HZ)) { + u8 status = readb(i2c->base + MPC_I2C_SR); + dev_dbg(i2c->dev, "timeout\n"); - if (readb(i2c->base + MPC_I2C_SR) == - (CSR_MCF | CSR_MBB | CSR_RXAK)) + if ((status & (CSR_MCF | CSR_MBB | CSR_RXAK)) != 0) { + writeb(status & ~CSR_MAL, + i2c->base + MPC_I2C_SR); mpc_i2c_fixup(i2c); + } return -EIO; } schedule();
This hunk looks good to me. It is basically what I did, except I didn't clear the MAL bit. Judging by the manual, it should be cleared.
quoted hunk ↗ jump to hunk
@@ -540,6 +566,14 @@ static int __devinit fsl_i2c_probe(struc } } + prop = of_get_property(op->node, "timeout", &plen); + if (prop && plen == sizeof(u32)) { + mpc_ops.timeout = *prop * HZ / 1000000; + if (mpc_ops.timeout < 5) + mpc_ops.timeout = 5; + } + dev_info(i2c->dev, "timeout %u us\n", mpc_ops.timeout * 1000000 / HZ); + dev_set_drvdata(&op->dev, i2c); i2c->adap = mpc_ops;
Ira