Skip to contentSkip to conversation

i2c: i2c-qcom-geni: More properly fix the DMA race

PATCH In review Douglas Anderson 5 participants 10 messages Oct 14, 2020 – Nov 3, 2020

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

v2 cover i2c: i2c-qcom-geni: More properly fix the DMA race

Previously I landed commit 02b9aec59243 ("i2c: i2c-qcom-geni: Fix DMA transfer race") to fix a race we were seeing. While that most definitely fixed the race we were seeing, it looks like it causes problems in the TX path, which we didn't stress test until we started trying to update firmware on devices.

Let's revert that patch and try another way: fix the original problem by disabling the interrupts that aren't relevant to DMA transfers. Now we can stress both TX and RX cases and see no problems. I also can't find any place to put an msleep() that causes problems anymore.

Since this problem only affects i2c, I'm hoping for an Ack from Bjorn and then all these patches can go through the i2c tree. However, if maintainers want to work a different way out to land that's OK too.

NOTE: the 3rd patch in the series could certianly be squashed with patch #1 or I could re-order / rejigger. To me it seemed like a good idea to first fix the probelm (and make the two functions as much of an inverse as possible) and later try to clean things up. Yell if you want something different.

Changes in v2:
- Consistently use "val_old" to keep track of old value.
- Add comments about why UART is special.

Douglas Anderson (3):
soc: qcom: geni: More properly switch to DMA mode Revert "i2c: i2c-qcom-geni: Fix DMA transfer race"
soc: qcom: geni: Optimize/comment select fifo/dma mode

 drivers/i2c/busses/i2c-qcom-geni.c |  6 ++--
 drivers/soc/qcom/qcom-geni-se.c    | 55 ++++++++++++++++++++++--------
 2 files changed, 43 insertions(+), 18 deletions(-)
Signature · 1 line
2.28.0.1011.ga647a8990f-goog
#1 lore raw

PATCH v2 1/3 soc: qcom: geni: More properly switch to DMA mode

On geni-i2c transfers using DMA, it was seen that if you program the command (I2C_READ) before calling geni_se_rx_dma_prep() that it could cause interrupts to fire. If we get unlucky, these interrupts can just keep firing (and not be handled) blocking further progress and hanging the system.

In commit 02b9aec59243 ("i2c: i2c-qcom-geni: Fix DMA transfer race") we avoided that by making sure we didn't program the command until after geni_se_rx_dma_prep() was called. While that avoided the problems, it also turns out to be invalid. At least in the TX case we started seeing sporadic corrupted transfers. This is easily seen by adding an msleep() between the DMA prep and the writing of the command, which makes the problem worse. That means we need to revert that commit and find another way to fix the bogus IRQs.

Specifically, after reverting commit 02b9aec59243 ("i2c:
i2c-qcom-geni: Fix DMA transfer race"), I put some traces in. I found that the when the interrupts were firing like crazy:
- "m_stat" had bits for M_RX_IRQ_EN, M_RX_FIFO_WATERMARK_EN set.
- "dma" was set.

Further debugging showed that I could make the problem happen more reliably by adding an "msleep(1)" any time after geni_se_setup_m_cmd() ran up until geni_se_rx_dma_prep() programmed the length.

A rather simple fix is to change geni_se_select_dma_mode() so it's a true inverse of geni_se_select_fifo_mode() and disables all the FIFO related interrupts. Now the problematic interrupts can't fire and we can program things in the correct order without worrying.

As part of this, let's also change the writel_relaxed() in the prepare function to a writel() so that our DMA is guaranteed to be prepared now that we can't rely on geni_se_setup_m_cmd()'s writel().

NOTE: the only current user of GENI_SE_DMA in mainline is i2c.

Fixes 37692de5d523 ("i2c: i2c-qcom-geni: Add bus driver for the Qualcomm GENI I2C controller")
Fixes 02b9aec59243 ("i2c: i2c-qcom-geni: Fix DMA transfer race")

---

(no changes since v1)

 drivers/soc/qcom/qcom-geni-se.c | 17 +++++++++++++++--
 1 file changed, 15 insertions(+), 2 deletions(-)
drivers/soc/qcom/qcom-geni-se.c +15 -2
index d0e4f520cff8..751a49f6534f 100644
@@ -289,10 +289,23 @@ static void geni_se_select_fifo_mode(struct geni_se *se)
 
 static void geni_se_select_dma_mode(struct geni_se *se)
 {
+	u32 proto = geni_se_read_proto(se);
 	u32 val;
 
 	geni_se_irq_clear(se);
 
+	val = readl_relaxed(se->base + SE_GENI_M_IRQ_EN);
+	if (proto != GENI_SE_UART) {
+		val &= ~(M_CMD_DONE_EN | M_TX_FIFO_WATERMARK_EN);
+		val &= ~(M_RX_FIFO_WATERMARK_EN | M_RX_FIFO_LAST_EN);
+	}
+	writel_relaxed(val, se->base + SE_GENI_M_IRQ_EN);
+
+	val = readl_relaxed(se->base + SE_GENI_S_IRQ_EN);
+	if (proto != GENI_SE_UART)
+		val &= ~S_CMD_DONE_EN;
+	writel_relaxed(val, se->base + SE_GENI_S_IRQ_EN);
+
 	val = readl_relaxed(se->base + SE_GENI_DMA_MODE_EN);
 	val |= GENI_DMA_MODE_EN;
 	writel_relaxed(val, se->base + SE_GENI_DMA_MODE_EN);
@@ -651,7 +664,7 @@ int geni_se_tx_dma_prep(struct geni_se *se, void *buf, size_t len,
 	writel_relaxed(lower_32_bits(*iova), se->base + SE_DMA_TX_PTR_L);
 	writel_relaxed(upper_32_bits(*iova), se->base + SE_DMA_TX_PTR_H);
 	writel_relaxed(GENI_SE_DMA_EOT_BUF, se->base + SE_DMA_TX_ATTR);
-	writel_relaxed(len, se->base + SE_DMA_TX_LEN);
+	writel(len, se->base + SE_DMA_TX_LEN);
 	return 0;
 }
 EXPORT_SYMBOL(geni_se_tx_dma_prep);
@@ -688,7 +701,7 @@ int geni_se_rx_dma_prep(struct geni_se *se, void *buf, size_t len,
 	writel_relaxed(upper_32_bits(*iova), se->base + SE_DMA_RX_PTR_H);
 	/* RX does not have EOT buffer type bit. So just reset RX_ATTR */
 	writel_relaxed(0, se->base + SE_DMA_RX_ATTR);
-	writel_relaxed(len, se->base + SE_DMA_RX_LEN);
+	writel(len, se->base + SE_DMA_RX_LEN);
 	return 0;
 }
 EXPORT_SYMBOL(geni_se_rx_dma_prep);
Signature · 1 line
2.28.0.1011.ga647a8990f-goog
#2 lore raw

PATCH v2 2/3 Revert "i2c: i2c-qcom-geni: Fix DMA transfer race"

In reply to Douglas Anderson

This reverts commit 02b9aec59243c6240fc42884acc958602146ddf6.

As talked about in the patch ("soc: qcom: geni: More properly switch to DMA mode"), swapping the order of geni_se_setup_m_cmd() and geni_se_xx_dma_prep() can sometimes cause corrupted transfers. Thus we traded one problem for another. Now that we've debugged the problem further and fixed the geni helper functions to more disable FIFO interrupts when we move to DMA mode we can revert it and end up with (hopefully) zero problems!

To be explicit, the patch ("soc: qcom: geni: More properly switch to DMA mode") is a prerequisite for this one.

Fixes 02b9aec59243 ("i2c: i2c-qcom-geni: Fix DMA transfer race")

---

(no changes since v1)

 drivers/i2c/busses/i2c-qcom-geni.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)
drivers/i2c/busses/i2c-qcom-geni.c +2 -4
index dead5db3315a..32b2a9921b14 100644
@@ -367,6 +367,7 @@ static int geni_i2c_rx_one_msg(struct geni_i2c_dev *gi2c, struct i2c_msg *msg,
 		geni_se_select_mode(se, GENI_SE_FIFO);
 
 	writel_relaxed(len, se->base + SE_I2C_RX_TRANS_LEN);
+	geni_se_setup_m_cmd(se, I2C_READ, m_param);
 
 	if (dma_buf && geni_se_rx_dma_prep(se, dma_buf, len, &rx_dma)) {
 		geni_se_select_mode(se, GENI_SE_FIFO);
@@ -374,8 +375,6 @@ static int geni_i2c_rx_one_msg(struct geni_i2c_dev *gi2c, struct i2c_msg *msg,
 		dma_buf = NULL;
 	}
 
-	geni_se_setup_m_cmd(se, I2C_READ, m_param);
-
 	time_left = wait_for_completion_timeout(&gi2c->done, XFER_TIMEOUT);
 	if (!time_left)
 		geni_i2c_abort_xfer(gi2c);
@@ -409,6 +408,7 @@ static int geni_i2c_tx_one_msg(struct geni_i2c_dev *gi2c, struct i2c_msg *msg,
 		geni_se_select_mode(se, GENI_SE_FIFO);
 
 	writel_relaxed(len, se->base + SE_I2C_TX_TRANS_LEN);
+	geni_se_setup_m_cmd(se, I2C_WRITE, m_param);
 
 	if (dma_buf && geni_se_tx_dma_prep(se, dma_buf, len, &tx_dma)) {
 		geni_se_select_mode(se, GENI_SE_FIFO);
@@ -416,8 +416,6 @@ static int geni_i2c_tx_one_msg(struct geni_i2c_dev *gi2c, struct i2c_msg *msg,
 		dma_buf = NULL;
 	}
 
-	geni_se_setup_m_cmd(se, I2C_WRITE, m_param);
-
 	if (!dma_buf) /* Get FIFO IRQ */
 		writel_relaxed(1, se->base + SE_GENI_TX_WATERMARK_REG);
 
Signature · 1 line
2.28.0.1011.ga647a8990f-goog
#3 lore raw

Re: [PATCH v2 2/3] Revert "i2c: i2c-qcom-geni: Fix DMA transfer race"

Quoted 10 lines · Douglas Anderson

This reverts commit 02b9aec59243c6240fc42884acc958602146ddf6.

As talked about in the patch ("soc: qcom: geni: More properly switch to DMA mode"), swapping the order of geni_se_setup_m_cmd() and geni_se_xx_dma_prep() can sometimes cause corrupted transfers. Thus we traded one problem for another. Now that we've debugged the problem further and fixed the geni helper functions to more disable FIFO interrupts when we move to DMA mode we can revert it and end up with (hopefully) zero problems!

Quoted 8 lines

To be explicit, the patch ("soc: qcom: geni: More properly switch to DMA mode") is a prerequisite for this one.

Fixes 02b9aec59243 ("i2c: i2c-qcom-geni: Fix DMA transfer race")

Wolfram, would you like to pick this patch or would you prefer that it goes together with the other two through the soc tree?

Per Doug's description it has a functional dependency on patch 1, but as long as all three patches makes it into v5.11 I'm okay with either way.

Regards,
Bjorn

Quoted 48 lines

---

(no changes since v1)

 drivers/i2c/busses/i2c-qcom-geni.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)
drivers/i2c/busses/i2c-qcom-geni.c +2 -4
index dead5db3315a..32b2a9921b14 100644
@@ -367,6 +367,7 @@ static int geni_i2c_rx_one_msg(struct geni_i2c_dev *gi2c, struct i2c_msg *msg,
 		geni_se_select_mode(se, GENI_SE_FIFO);
 
 	writel_relaxed(len, se->base + SE_I2C_RX_TRANS_LEN);
+	geni_se_setup_m_cmd(se, I2C_READ, m_param);
 
 	if (dma_buf && geni_se_rx_dma_prep(se, dma_buf, len, &rx_dma)) {
 		geni_se_select_mode(se, GENI_SE_FIFO);
@@ -374,8 +375,6 @@ static int geni_i2c_rx_one_msg(struct geni_i2c_dev *gi2c, struct i2c_msg *msg,
 		dma_buf = NULL;
 	}
 
-	geni_se_setup_m_cmd(se, I2C_READ, m_param);
-
 	time_left = wait_for_completion_timeout(&gi2c->done, XFER_TIMEOUT);
 	if (!time_left)
 		geni_i2c_abort_xfer(gi2c);
@@ -409,6 +408,7 @@ static int geni_i2c_tx_one_msg(struct geni_i2c_dev *gi2c, struct i2c_msg *msg,
 		geni_se_select_mode(se, GENI_SE_FIFO);
 
 	writel_relaxed(len, se->base + SE_I2C_TX_TRANS_LEN);
+	geni_se_setup_m_cmd(se, I2C_WRITE, m_param);
 
 	if (dma_buf && geni_se_tx_dma_prep(se, dma_buf, len, &tx_dma)) {
 		geni_se_select_mode(se, GENI_SE_FIFO);
@@ -416,8 +416,6 @@ static int geni_i2c_tx_one_msg(struct geni_i2c_dev *gi2c, struct i2c_msg *msg,
 		dma_buf = NULL;
 	}
 
-	geni_se_setup_m_cmd(se, I2C_WRITE, m_param);
-
 	if (!dma_buf) /* Get FIFO IRQ */
 		writel_relaxed(1, se->base + SE_GENI_TX_WATERMARK_REG);
 
Signature · 1 line
2.28.0.1011.ga647a8990f-goog
#4 lore raw

Re: [PATCH v2 2/3] Revert "i2c: i2c-qcom-geni: Fix DMA transfer race"

Quoted 3 lines

Wolfram, would you like to pick this patch or would you prefer that it goes together with the other two through the soc tree?

Actually, I prefer the soc tree because of the functional dependency. I am not aware of any pending qcom-geni patches, yet I think an immutable branch for me to pull in would be nice in this case. Could you provide one for me?

#5 lore raw

Re: [PATCH v2 2/3] Revert "i2c: i2c-qcom-geni: Fix DMA transfer race"

Quoted 9 lines · Wolfram Sang
Quoted 3 lines

Wolfram, would you like to pick this patch or would you prefer that it goes together with the other two through the soc tree?

Actually, I prefer the soc tree because of the functional dependency. I am not aware of any pending qcom-geni patches, yet I think an immutable branch for me to pull in would be nice in this case. Could you provide one for me?

Sounds good, please find the series applied on top of v5.10-rc1 at:

https://git.kernel.org/pub/scm/linux/kernel/git/qcom/linux.git tags/[email protected]

I've merged the same into the qcom tree for 5.11.

Regards,
Bjorn

#7 lore raw

PATCH v2 3/3 soc: qcom: geni: Optimize/comment select fifo/dma mode

In reply to Douglas Anderson

The functions geni_se_select_fifo_mode() and geni_se_select_fifo_mode() are a little funny. They read/write a bunch of memory mapped registers even if they don't change or aren't relevant for the current protocol. Let's make them a little more sane. We'll also add a comment explaining why we don't do some of the operations for UART.

NOTE: there is no evidence at all that this makes any performance difference and it fixes no bugs. However, it seems (to me) like it makes the functions a little easier to understand. Decreasing the amount of times we read/write memory mapped registers is also nice, even if we are using "relaxed" variants.

---
Note that I didn't add Dmitry Baryshkov's Tested-by tag to the 3rd patch since it changed subtly. Also note that when adding comments about why the UART is special it seems clear to me that we really shouldn't be managing these interrupt enables in the common code. It seems like drivers should manage / enable the interrupts that they care about. That seems like a bigger change, though, and I didn't want to muddy the waters and potentially delay the important fix from patch #1 and #2. Hopefully someone from Qualcomm can take on cleaning this stuff up after these fixes land.

Changes in v2:
- Consistently use "val_old" to keep track of old value.
- Add comments about why UART is special.

 drivers/soc/qcom/qcom-geni-se.c | 50 +++++++++++++++++++++------------
 1 file changed, 32 insertions(+), 18 deletions(-)
drivers/soc/qcom/qcom-geni-se.c +32 -18
index 751a49f6534f..7649b2057b9a 100644
@@ -266,49 +266,63 @@ EXPORT_SYMBOL(geni_se_init);
 static void geni_se_select_fifo_mode(struct geni_se *se)
 {
 	u32 proto = geni_se_read_proto(se);
-	u32 val;
+	u32 val, val_old;
 
 	geni_se_irq_clear(se);
 
-	val = readl_relaxed(se->base + SE_GENI_M_IRQ_EN);
+	/*
+	 * The RX path for the UART is asynchronous and so needs more
+	 * complex logic for enabling / disabling its interrupts.
+	 *
+	 * Specific notes:
+	 * - The done and TX-related interrupts are managed manually.
+	 * - We don't RX from the main sequencer (we use the secondary) so
+	 *   we don't need the RX-related interrupts enabled in the main
+	 *   sequencer for UART.
+	 */
 	if (proto != GENI_SE_UART) {
+		val_old = val = readl_relaxed(se->base + SE_GENI_M_IRQ_EN);
 		val |= M_CMD_DONE_EN | M_TX_FIFO_WATERMARK_EN;
 		val |= M_RX_FIFO_WATERMARK_EN | M_RX_FIFO_LAST_EN;
-	}
-	writel_relaxed(val, se->base + SE_GENI_M_IRQ_EN);
+		if (val != val_old)
+			writel_relaxed(val, se->base + SE_GENI_M_IRQ_EN);
 
-	val = readl_relaxed(se->base + SE_GENI_S_IRQ_EN);
-	if (proto != GENI_SE_UART)
+		val_old = val = readl_relaxed(se->base + SE_GENI_S_IRQ_EN);
 		val |= S_CMD_DONE_EN;
-	writel_relaxed(val, se->base + SE_GENI_S_IRQ_EN);
+		if (val != val_old)
+			writel_relaxed(val, se->base + SE_GENI_S_IRQ_EN);
+	}
 
-	val = readl_relaxed(se->base + SE_GENI_DMA_MODE_EN);
+	val_old = val = readl_relaxed(se->base + SE_GENI_DMA_MODE_EN);
 	val &= ~GENI_DMA_MODE_EN;
-	writel_relaxed(val, se->base + SE_GENI_DMA_MODE_EN);
+	if (val != val_old)
+		writel_relaxed(val, se->base + SE_GENI_DMA_MODE_EN);
 }
 
 static void geni_se_select_dma_mode(struct geni_se *se)
 {
 	u32 proto = geni_se_read_proto(se);
-	u32 val;
+	u32 val, val_old;
 
 	geni_se_irq_clear(se);
 
-	val = readl_relaxed(se->base + SE_GENI_M_IRQ_EN);
 	if (proto != GENI_SE_UART) {
+		val_old = val = readl_relaxed(se->base + SE_GENI_M_IRQ_EN);
 		val &= ~(M_CMD_DONE_EN | M_TX_FIFO_WATERMARK_EN);
 		val &= ~(M_RX_FIFO_WATERMARK_EN | M_RX_FIFO_LAST_EN);
-	}
-	writel_relaxed(val, se->base + SE_GENI_M_IRQ_EN);
+		if (val != val_old)
+			writel_relaxed(val, se->base + SE_GENI_M_IRQ_EN);
 
-	val = readl_relaxed(se->base + SE_GENI_S_IRQ_EN);
-	if (proto != GENI_SE_UART)
+		val_old = val = readl_relaxed(se->base + SE_GENI_S_IRQ_EN);
 		val &= ~S_CMD_DONE_EN;
-	writel_relaxed(val, se->base + SE_GENI_S_IRQ_EN);
+		if (val != val_old)
+			writel_relaxed(val, se->base + SE_GENI_S_IRQ_EN);
+	}
 
-	val = readl_relaxed(se->base + SE_GENI_DMA_MODE_EN);
+	val_old = val = readl_relaxed(se->base + SE_GENI_DMA_MODE_EN);
 	val |= GENI_DMA_MODE_EN;
-	writel_relaxed(val, se->base + SE_GENI_DMA_MODE_EN);
+	if (val != val_old)
+		writel_relaxed(val, se->base + SE_GENI_DMA_MODE_EN);
 }
 
 /**
Signature · 1 line
2.28.0.1011.ga647a8990f-goog
Reviewed-by on v2 3/3
#8 lore raw

Re: [PATCH v2 3/3] soc: qcom: geni: Optimize/comment select fifo/dma mode

Quoted 7 lines · Douglas Anderson

The functions geni_se_select_fifo_mode() and geni_se_select_fifo_mode() are a little funny. They read/write a bunch of memory mapped registers even if they don't change or aren't relevant for the current protocol. Let's make them a little more sane. We'll also add a comment explaining why we don't do some of the operations for UART.

Quoted 7 lines

NOTE: there is no evidence at all that this makes any performance difference and it fixes no bugs. However, it seems (to me) like it makes the functions a little easier to understand. Decreasing the amount of times we read/write memory mapped registers is also nice, even if we are using "relaxed" variants.

The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,\na Linux Foundation Collaborative Project

Reviewed-by on v2 3/3
#9 lore raw

Re: [PATCH v2 3/3] soc: qcom: geni: Optimize/comment select fifo/dma mode

In reply to Douglas Anderson

Quoting Douglas Anderson (2020-10-13 14:25:30)

Quoted 7 lines

The functions geni_se_select_fifo_mode() and geni_se_select_fifo_mode() are a little funny. They read/write a bunch of memory mapped registers even if they don't change or aren't relevant for the current protocol. Let's make them a little more sane. We'll also add a comment explaining why we don't do some of the operations for UART.

Quoted 8 lines

NOTE: there is no evidence at all that this makes any performance difference and it fixes no bugs. However, it seems (to me) like it makes the functions a little easier to understand. Decreasing the amount of times we read/write memory mapped registers is also nice, even if we are using "relaxed" variants.

---