[PATCH v4 0/3] i2c: qcom-geni: Add I2C bus recovery support

[PATCH v4 0/3] i2c: qcom-geni: Add I2C bus recovery support

[Aniket Randive]

An interrupted I2C transaction can leave a target holding SDA low and
prevent subsequent transfers. Add recovery using the GENI I2C_BUS_CLEAR
and I2C_STOP_ON_BUS opcodes in both FIFO and GPI DMA modes, without
requiring GPIO-based recovery.

The first patch adds zero-payload recovery opcode support to the GPI DMA
driver. The second registers the I2C bus recovery callback and implements
FIFO recovery. The third extends recovery to GPI DMA mode.

Recovery checks both SDA and SCL, skips an idle bus, and rejects recovery
when SCL is held low. It is available through i2c_recover_bus() and runs
automatically after relevant transfer errors, preserving the original
transfer error. GPI recovery stops after a DMA command timeout.

Signed-off-by: Aniket Randive <aniket.randive@oss.qualcomm.com>
---
Changes in v4:
- Clarify that gpi_prep_slave_sg() is common to all protocols.
- Consolidate runtime-PM cleanup at out_pm.
- Use a ternary expression for the recovery result.
- Preserve DMA submission errors.
- Move reinit_completion() before dmaengine_submit().
- Link to v3: https://patch.msgid.link/20261027-20260826_aniket_randive_oss_qualcomm_com-v3-0-7f2a6e572de9@oss.qualcomm.com

Changes in v3:
- Use the adapter timeout from the FIFO recovery patch onward.
- Configure clocks and timing before recovery, with PM cleanup on failure.
- Remove unreachable GPI error checks.
- Log DMA termination failures during GPI recovery.
- Preserve GPI timeouts and stop further recovery commands.
- Link to v2: https://patch.msgid.link/20261023-20260826_aniket_randive_oss_qualcomm_com-v2-0-93572b453d6f@oss.qualcomm.com

Changes in v2:
- Rebase onto next-20261006.
- Move the recovery opcode macro removal into the DMA opcode patch.
- Check both SDA and SCL for bus idle and reject recovery with SCL low.
- Share the bus-state and recovery-trigger checks between both modes.
- Guard the current message pointer when checking data NACK status.
- Reuse the FIFO transfer abort helper for recovery command timeouts.
- Use the adapter timeout for FIFO and GPI recovery commands.
- Preserve the DMA configuration error code in GPI recovery.
- Link to v1: https://lore.kernel.org/r/20260826-i2c_bus_recovery-v1-0-203f0ec76f84@oss.qualcomm.com

---
Aniket Randive (3):
      dmaengine: qcom: gpi: Add I2C bus recovery opcode support
      i2c: qcom-geni: Add bus recovery support for FIFO mode
      i2c: qcom-geni: Add bus recovery support for GPI DMA mode

 drivers/dma/qcom/gpi.c             |  30 ++++++
 drivers/i2c/busses/i2c-qcom-geni.c | 207 ++++++++++++++++++++++++++++++++++++-
 include/linux/dma/qcom-gpi-dma.h   |   2 +
 3 files changed, 236 insertions(+), 3 deletions(-)
---
base-commit: eea3fef32a9cf36abcb5975a5a594e4135a6b026
change-id: 20261008-20260826_aniket_randive_oss_qualcomm_com-84f5f536302d

Best regards,
--  
Aniket Randive <aniket.randive@oss.qualcomm.com>

[PATCH v4 1/3] dmaengine: qcom: gpi: Add I2C bus recovery opcode support

[Aniket Randive]

The I2C_BUS_CLEAR and I2C_STOP_ON_BUS GENI sequencer opcodes release a
stuck I2C bus, but the GPI DMA driver has no way to emit these
zero-payload control commands, so an I2C controller using GPI DMA
cannot recover the bus.

Add I2C_BUS_CLEAR and I2C_STOP_ON_BUS to enum i2c_op so I2C drivers can
request them through gpi_i2c_config.op. Handle them in
gpi_create_i2c_tre() by emitting a CONFIG TRE (when set_config is set)
followed by a GO TRE with IEOT set and no DMA TRE, and reserve the
matching number of ring entries in gpi_prep_slave_sg().

Signed-off-by: Aniket Randive <aniket.randive@oss.qualcomm.com>
---
 drivers/dma/qcom/gpi.c             | 30 ++++++++++++++++++++++++++++++
 drivers/i2c/busses/i2c-qcom-geni.c |  2 --
 include/linux/dma/qcom-gpi-dma.h   |  2 ++
 3 files changed, 32 insertions(+), 2 deletions(-)

diff --git a/drivers/dma/qcom/gpi.c b/drivers/dma/qcom/gpi.c
index c41bfac65ddf..3dda58e6a99d 100644
--- a/drivers/dma/qcom/gpi.c
+++ b/drivers/dma/qcom/gpi.c
@@ -1620,6 +1620,17 @@ gpi_peripheral_config(struct dma_chan *chan, struct dma_slave_config *config)
 	return 0;
 }
 
+static void gpi_create_i2c_go_recovery_tre(struct gpi_i2c_config *i2c,
+					   struct gpi_tre *tre)
+{
+	tre->dword[0] = u32_encode_bits(i2c->op, TRE_I2C_GO_CMD);
+	tre->dword[1] = 0;
+	tre->dword[2] = 0;
+
+	tre->dword[3] = u32_encode_bits(TRE_TYPE_GO, TRE_FLAGS_TYPE);
+	tre->dword[3] |= u32_encode_bits(1, TRE_FLAGS_IEOT);
+}
+
 static int gpi_create_i2c_tre(struct gchan *chan, struct gpi_desc *desc,
 			      struct scatterlist *sgl, enum dma_transfer_direction direction,
 			      unsigned long flags)
@@ -1650,6 +1661,12 @@ static int gpi_create_i2c_tre(struct gchan *chan, struct gpi_desc *desc,
 		tre->dword[3] |= u32_encode_bits(1, TRE_FLAGS_CHAIN);
 	}
 
+	if (i2c->op == I2C_BUS_CLEAR || i2c->op == I2C_STOP_ON_BUS) {
+		gpi_create_i2c_go_recovery_tre(i2c, &desc->tre[tre_idx]);
+		tre_idx++;
+		goto log_tre;
+	}
+
 	/* create the GO tre for Tx */
 	if (i2c->op == I2C_WRITE) {
 		tre = &desc->tre[tre_idx];
@@ -1692,6 +1709,7 @@ static int gpi_create_i2c_tre(struct gchan *chan, struct gpi_desc *desc,
 			tre->dword[3] |= u32_encode_bits(1, TRE_FLAGS_BEI);
 	}
 
+log_tre:
 	for (i = 0; i < tre_idx; i++)
 		dev_dbg(dev, "TRE:%d %x:%x:%x:%x\n", i, desc->tre[i].dword[0],
 			desc->tre[i].dword[1], desc->tre[i].dword[2], desc->tre[i].dword[3]);
@@ -1825,6 +1843,18 @@ gpi_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl,
 		nr_tre = 2;
 	if (direction == DMA_DEV_TO_MEM) /* rx */
 		nr_tre = 1;
+	/*
+	 * Recovery opcodes do not require DMA data TREs, only CONFIG
+	 * (for set_config) and GO TREs. Since gpi_prep_slave_sg() is
+	 * common to all protocols, verify the channel is I2C before
+	 * accessing the configuration data.
+	 */
+	if (gchan->protocol == QCOM_GPI_I2C) {
+		struct gpi_i2c_config *i2c = gchan->config;
+
+		if (i2c->op == I2C_BUS_CLEAR || i2c->op == I2C_STOP_ON_BUS)
+			nr_tre = set_config ? 2 : 1;
+	}
 
 	/* calculate # of elements required & available */
 	nr = gpi_ring_num_elements_avail(ch_ring);
diff --git a/drivers/i2c/busses/i2c-qcom-geni.c b/drivers/i2c/busses/i2c-qcom-geni.c
index 5cc47b670bd1..60b6273f538c 100644
--- a/drivers/i2c/busses/i2c-qcom-geni.c
+++ b/drivers/i2c/busses/i2c-qcom-geni.c
@@ -34,8 +34,6 @@
 #define I2C_READ		0x2
 #define I2C_WRITE_READ		0x3
 #define I2C_ADDR_ONLY		0x4
-#define I2C_BUS_CLEAR		0x6
-#define I2C_STOP_ON_BUS		0x7
 /* M_CMD params for I2C */
 #define PRE_CMD_DELAY		BIT(0)
 #define TIMESTAMP_BEFORE	BIT(1)
diff --git a/include/linux/dma/qcom-gpi-dma.h b/include/linux/dma/qcom-gpi-dma.h
index 332be28427e4..b9b75c302825 100644
--- a/include/linux/dma/qcom-gpi-dma.h
+++ b/include/linux/dma/qcom-gpi-dma.h
@@ -52,6 +52,8 @@ struct gpi_spi_config {
 enum i2c_op {
 	I2C_WRITE = 1,
 	I2C_READ,
+	I2C_BUS_CLEAR = 6,
+	I2C_STOP_ON_BUS,
 };
 
 /**

-- 
2.34.1

Re: [PATCH v4 1/3] dmaengine: qcom: gpi: Add I2C bus recovery opcode support

[Mukesh Savaliya]

On 11/3/2026 2:26 PM, Aniket Randive wrote:
> The I2C_BUS_CLEAR and I2C_STOP_ON_BUS GENI sequencer opcodes release a
> stuck I2C bus, but the GPI DMA driver has no way to emit these
> zero-payload control commands, so an I2C controller using GPI DMA
> cannot recover the bus.
> 
> Add I2C_BUS_CLEAR and I2C_STOP_ON_BUS to enum i2c_op so I2C drivers can
> request them through gpi_i2c_config.op. Handle them in
> gpi_create_i2c_tre() by emitting a CONFIG TRE (when set_config is set)
> followed by a GO TRE with IEOT set and no DMA TRE, and reserve the
> matching number of ring entries in gpi_prep_slave_sg().
> 
> Signed-off-by: Aniket Randive <aniket.randive@oss.qualcomm.com>
> ---
Reviewed-by: Mukesh Kumar Savaliya <mukesh.savaliya@oss.qualcomm.com>

[PATCH v4 2/3] i2c: qcom-geni: Add bus recovery support for FIFO mode

[Aniket Randive]

I2C transfers in FIFO mode can fail with -EPROTO, -ETIMEDOUT or
-EAGAIN when a target holds SDA low, leaving the bus stuck and
preventing subsequent transactions.

Add bus recovery support using the I2C_BUS_CLEAR and I2C_STOP_ON_BUS
hardware opcodes to restore the bus to an idle state. Check the SDA
line state through SE_GENI_IOS.RX_DATA_IN and skip recovery when the
bus is already free. Trigger recovery automatically from
geni_i2c_xfer() on -EPROTO, -ETIMEDOUT and -EAGAIN (arbitration-lost)
errors, and register the recovery callback through i2c_bus_recovery_info
to allow recovery via i2c_recover_bus().

Use the adapter timeout for recovery commands. Configure the serial clock
and bus timing after runtime resume so direct calls to i2c_recover_bus()
do not depend on an earlier transfer having configured them.

This adds bus recovery support for FIFO mode only.

Signed-off-by: Aniket Randive <aniket.randive@oss.qualcomm.com>
---
 drivers/i2c/busses/i2c-qcom-geni.c | 124 ++++++++++++++++++++++++++++++++++++-
 1 file changed, 123 insertions(+), 1 deletion(-)

diff --git a/drivers/i2c/busses/i2c-qcom-geni.c b/drivers/i2c/busses/i2c-qcom-geni.c
index 60b6273f538c..b08353616adc 100644
--- a/drivers/i2c/busses/i2c-qcom-geni.c
+++ b/drivers/i2c/busses/i2c-qcom-geni.c
@@ -34,6 +34,7 @@
 #define I2C_READ		0x2
 #define I2C_WRITE_READ		0x3
 #define I2C_ADDR_ONLY		0x4
+#define I2C_BUS_FREE		(RX_DATA_IN | IO2_DATA_IN)
 /* M_CMD params for I2C */
 #define PRE_CMD_DELAY		BIT(0)
 #define TIMESTAMP_BEFORE	BIT(1)
@@ -145,6 +146,7 @@ struct geni_i2c_dev {
 	u32 num_msgs;
 	struct geni_i2c_gpi_multi_desc_xfer i2c_multi_desc_config;
 	const struct geni_i2c_desc *dev_data;
+	struct i2c_bus_recovery_info rinfo;
 };
 
 struct geni_i2c_err_log {
@@ -316,7 +318,7 @@ static void geni_i2c_check_addr_data_nack(struct geni_i2c_dev *gi2c)
 {
 	if (!readl_relaxed(gi2c->se.base + SE_GENI_M_GP_LENGTH))
 		geni_i2c_err(gi2c, ADDR_NACK);
-	else if (!(gi2c->cur->flags & I2C_M_RD))
+	else if (gi2c->cur && !(gi2c->cur->flags & I2C_M_RD))
 		geni_i2c_err(gi2c, DATA_NACK);
 }
 
@@ -1013,6 +1015,112 @@ static int geni_i2c_fifo_xfer(struct geni_i2c_dev *gi2c,
 	return num;
 }
 
+static int geni_i2c_fifo_bus_cmd(struct geni_i2c_dev *gi2c, u32 cmd)
+{
+	unsigned long time_left;
+	unsigned long flags;
+
+	/*
+	 * Clear cur so the IRQ handler does not attempt FIFO watermark
+	 * filling or draining while the recovery opcode is in flight.
+	 * cur and err are shared with geni_i2c_irq(), which reads cur and
+	 * writes err under gi2c->lock, so take the lock around this reset.
+	 */
+	spin_lock_irqsave(&gi2c->lock, flags);
+	gi2c->cur = NULL;
+	gi2c->err = 0;
+	spin_unlock_irqrestore(&gi2c->lock, flags);
+	geni_se_select_mode(&gi2c->se, GENI_SE_FIFO);
+	reinit_completion(&gi2c->done);
+
+	geni_se_setup_m_cmd(&gi2c->se, cmd, 0);
+	time_left = wait_for_completion_timeout(&gi2c->done, gi2c->adap.timeout);
+	if (!time_left) {
+		dev_err(gi2c->se.dev, "timeout waiting for bus cmd %u\n", cmd);
+		geni_i2c_abort_xfer(gi2c);
+
+		return -ETIMEDOUT;
+	}
+
+	/*
+	 * ARB_LOST and BUS_PROTO interrupts may be reported while the bus
+	 * transitions from stuck to idle during the recovery sequence.
+	 * The opcode completed successfully so treat these as success.
+	 */
+	if (gi2c->err == -EAGAIN || gi2c->err == -EPROTO)
+		return 0;
+
+	return gi2c->err;
+}
+
+static bool geni_i2c_bus_is_free(struct geni_i2c_dev *gi2c)
+{
+	return (readl_relaxed(gi2c->se.base + SE_GENI_IOS) & I2C_BUS_FREE) ==
+	       I2C_BUS_FREE;
+}
+
+static bool geni_i2c_needs_bus_recovery(struct geni_i2c_dev *gi2c, int ret)
+{
+	/*
+	 * -EPROTO and -ETIMEDOUT indicate bus errors where the target may be
+	 * holding a bus line low. ARB_LOST (-EAGAIN) on a single-controller
+	 * bus indicates a stuck target, not a real arbitration loss.
+	 * GPI DMA mode extends this trigger separately.
+	 */
+	return !gi2c->gpi_mode &&
+	       (ret == -EPROTO || ret == -ETIMEDOUT || ret == -EAGAIN) &&
+	       !geni_i2c_bus_is_free(gi2c);
+}
+
+static int geni_i2c_recover_bus(struct i2c_adapter *adap)
+{
+	struct geni_i2c_dev *gi2c = i2c_get_adapdata(adap);
+	u32 ios;
+	int ret;
+
+	ret = pm_runtime_get_sync(gi2c->se.dev);
+	if (ret < 0) {
+		dev_err(gi2c->se.dev, "bus recovery failed, error turning SE resources:%d\n", ret);
+		pm_runtime_put_noidle(gi2c->se.dev);
+		return ret;
+	}
+
+	ret = gi2c->dev_data->set_rate(&gi2c->se, gi2c->clk_freq_out);
+	if (ret) {
+		dev_err(gi2c->se.dev, "bus recovery failed, error setting clock rate: %d\n",
+			ret);
+		goto out_pm;
+	}
+
+	if (geni_i2c_bus_is_free(gi2c)) {
+		ret = 0;
+		goto out_pm;
+	}
+
+	/* Bus clear cannot recover a target holding SCL low. */
+	ios = readl_relaxed(gi2c->se.base + SE_GENI_IOS);
+	if (!(ios & IO2_DATA_IN)) {
+		ret = -EBUSY;
+		goto out_pm;
+	}
+
+	ret = geni_i2c_fifo_bus_cmd(gi2c, I2C_BUS_CLEAR);
+	if (!ret)
+		ret = geni_i2c_fifo_bus_cmd(gi2c, I2C_STOP_ON_BUS);
+
+	/*
+	 * Recovery succeeds only once both bus lines are released. The bus
+	 * state is the authority, while a clean opcode status with either
+	 * line low is a failed recovery.
+	 */
+	ret = geni_i2c_bus_is_free(gi2c) ? 0 : (ret ?: -EBUSY);
+
+out_pm:
+	pm_runtime_put_autosuspend(gi2c->se.dev);
+
+	return ret;
+}
+
 static int geni_i2c_xfer(struct i2c_adapter *adap,
 			 struct i2c_msg msgs[],
 			 int num)
@@ -1037,6 +1145,15 @@ static int geni_i2c_xfer(struct i2c_adapter *adap,
 	else
 		ret = geni_i2c_fifo_xfer(gi2c, msgs, num);
 
+	if (geni_i2c_needs_bus_recovery(gi2c, ret)) {
+		int recovery_ret = i2c_recover_bus(adap);
+
+		if (recovery_ret)
+			dev_err(gi2c->se.dev,
+				"bus recovery failed: %d (xfer error: %d)\n",
+				recovery_ret, ret);
+	}
+
 out_pm:
 	pm_runtime_put_autosuspend(gi2c->se.dev);
 	gi2c->cur = NULL;
@@ -1239,6 +1356,11 @@ static int geni_i2c_probe(struct platform_device *pdev)
 	if (ret < 0)
 		return ret;
 
+	if (!gi2c->gpi_mode) {
+		gi2c->rinfo.recover_bus = geni_i2c_recover_bus;
+		gi2c->adap.bus_recovery_info = &gi2c->rinfo;
+	}
+
 	ret = i2c_add_adapter(&gi2c->adap);
 	if (ret) {
 		release_gpi_dma(gi2c);

-- 
2.34.1

Re: [PATCH v4 2/3] i2c: qcom-geni: Add bus recovery support for FIFO mode

[Mukesh Savaliya]

On 11/3/2026 2:26 PM, Aniket Randive wrote:
> I2C transfers in FIFO mode can fail with -EPROTO, -ETIMEDOUT or
> -EAGAIN when a target holds SDA low, leaving the bus stuck and
> preventing subsequent transactions.
> 
> Add bus recovery support using the I2C_BUS_CLEAR and I2C_STOP_ON_BUS
> hardware opcodes to restore the bus to an idle state. Check the SDA
> line state through SE_GENI_IOS.RX_DATA_IN and skip recovery when the
> bus is already free. Trigger recovery automatically from
> geni_i2c_xfer() on -EPROTO, -ETIMEDOUT and -EAGAIN (arbitration-lost)
> errors, and register the recovery callback through i2c_bus_recovery_info
> to allow recovery via i2c_recover_bus().
> 
> Use the adapter timeout for recovery commands. Configure the serial clock
> and bus timing after runtime resume so direct calls to i2c_recover_bus()
> do not depend on an earlier transfer having configured them.
> 
> This adds bus recovery support for FIFO mode only.
> 
> Signed-off-by: Aniket Randive <aniket.randive@oss.qualcomm.com>
> ---
Reviewed-by: Mukesh Kumar Savaliya <mukesh.savaliya@oss.qualcomm.com>

[PATCH v4 3/3] i2c: qcom-geni: Add bus recovery support for GPI DMA mode

[Aniket Randive]

I2C transfers in GPI DMA mode can fail with -EIO when a target holds
SDA low, leaving the bus stuck and preventing subsequent transactions.
The GPI completion callback reports a NACK and a genuine bus error
alike as -EIO, and the existing recovery path switches the SE to FIFO
mode which is not valid in GPI DMA mode.

Add bus recovery support using the I2C_BUS_CLEAR and I2C_STOP_ON_BUS
opcodes issued via the GPI TX DMA channel. Guard the -EIO recovery
trigger with a SE_GENI_IOS I2C_BUS_FREE check to skip recovery when both
SDA and SCL are already released. Extend the recovery callback to dispatch
through the GPI DMA path when gi2c->gpi_mode is set.

Use the adapter timeout for GPI recovery commands, and preserve the DMA
configuration error code when GPI setup fails.

Stop recovery on a GPI command timeout and report DMA termination
failures. Do not issue another command or override the timeout with a
bus-idle result, since DMA cleanup may not have completed successfully.

Signed-off-by: Aniket Randive <aniket.randive@oss.qualcomm.com>
---
 drivers/i2c/busses/i2c-qcom-geni.c | 107 ++++++++++++++++++++++++++++++++-----
 1 file changed, 94 insertions(+), 13 deletions(-)

diff --git a/drivers/i2c/busses/i2c-qcom-geni.c b/drivers/i2c/busses/i2c-qcom-geni.c
index b08353616adc..d5d5d90544e8 100644
--- a/drivers/i2c/busses/i2c-qcom-geni.c
+++ b/drivers/i2c/busses/i2c-qcom-geni.c
@@ -1065,13 +1065,82 @@ static bool geni_i2c_needs_bus_recovery(struct geni_i2c_dev *gi2c, int ret)
 	 * -EPROTO and -ETIMEDOUT indicate bus errors where the target may be
 	 * holding a bus line low. ARB_LOST (-EAGAIN) on a single-controller
 	 * bus indicates a stuck target, not a real arbitration loss.
-	 * GPI DMA mode extends this trigger separately.
+	 * In GPI DMA mode the completion callback reports NACK, BUS_PROTO and
+	 * ARB_LOST alike as -EIO, so the SE_GENI_IOS I2C_BUS_FREE check
+	 * distinguishes a genuinely stuck bus from a benign NACK. Recovery is
+	 * skipped when both SDA and SCL are released (high) after the STOP.
 	 */
-	return !gi2c->gpi_mode &&
-	       (ret == -EPROTO || ret == -ETIMEDOUT || ret == -EAGAIN) &&
+	return (ret == -EPROTO || ret == -ETIMEDOUT || ret == -EAGAIN ||
+		(gi2c->gpi_mode && ret == -EIO)) &&
 	       !geni_i2c_bus_is_free(gi2c);
 }
 
+static int geni_i2c_gpi_bus_cmd(struct geni_i2c_dev *gi2c, enum i2c_op cmd)
+{
+	const struct geni_i2c_clk_fld *itr = gi2c->clk_fld;
+	struct dma_async_tx_descriptor *desc;
+	struct gpi_i2c_config peripheral = {};
+	struct dma_slave_config config = {};
+	unsigned long time_left;
+	unsigned long flags;
+	dma_cookie_t cookie;
+	int ret;
+
+	config.peripheral_config = &peripheral;
+	config.peripheral_size = sizeof(peripheral);
+
+	peripheral.set_config  = 1;
+	peripheral.pack_enable = I2C_PACK_TX | I2C_PACK_RX;
+	peripheral.cycle_count = itr->t_cycle_cnt;
+	peripheral.high_count  = itr->t_high_cnt;
+	peripheral.low_count   = itr->t_low_cnt;
+	peripheral.clk_div     = itr->clk_div;
+	peripheral.op          = cmd;
+
+	ret = dmaengine_slave_config(gi2c->tx_c, &config);
+	if (ret) {
+		dev_err(gi2c->se.dev, "dma config error %d for bus cmd %u\n", ret, cmd);
+		return ret;
+	}
+
+	desc = dmaengine_prep_slave_single(gi2c->tx_c, 0, 0, DMA_MEM_TO_DEV,
+					   DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
+	if (!desc)
+		return -EIO;
+
+	desc->callback_result = i2c_gpi_cb_result;
+	desc->callback_param = gi2c;
+
+	reinit_completion(&gi2c->done);
+	cookie = dmaengine_submit(desc);
+	ret = dma_submit_error(cookie);
+	if (ret)
+		return ret;
+
+	/*
+	 * cur and err are shared with geni_i2c_irq() and the GPI callback
+	 * i2c_gpi_cb_result(); both write err (the IRQ handler under
+	 * gi2c->lock). Reset them under the lock before issuing the transfer.
+	 */
+	spin_lock_irqsave(&gi2c->lock, flags);
+	gi2c->cur = NULL;
+	gi2c->err = 0;
+	spin_unlock_irqrestore(&gi2c->lock, flags);
+	dma_async_issue_pending(gi2c->tx_c);
+
+	time_left = wait_for_completion_timeout(&gi2c->done, gi2c->adap.timeout);
+	if (!time_left) {
+		dev_err(gi2c->se.dev, "timeout waiting for GPI bus cmd %u\n", cmd);
+		ret = dmaengine_terminate_sync(gi2c->tx_c);
+		if (ret)
+			dev_err(gi2c->se.dev, "DMA termination failed for bus cmd %u: %d\n",
+				cmd, ret);
+		return -ETIMEDOUT;
+	}
+
+	return gi2c->err;
+}
+
 static int geni_i2c_recover_bus(struct i2c_adapter *adap)
 {
 	struct geni_i2c_dev *gi2c = i2c_get_adapdata(adap);
@@ -1104,14 +1173,28 @@ static int geni_i2c_recover_bus(struct i2c_adapter *adap)
 		goto out_pm;
 	}
 
-	ret = geni_i2c_fifo_bus_cmd(gi2c, I2C_BUS_CLEAR);
-	if (!ret)
-		ret = geni_i2c_fifo_bus_cmd(gi2c, I2C_STOP_ON_BUS);
+	if (gi2c->gpi_mode) {
+		ret = geni_i2c_gpi_bus_cmd(gi2c, I2C_BUS_CLEAR);
+		/* Do not issue more commands after a timeout: DMA cleanup may have failed. */
+		if (ret == -ETIMEDOUT)
+			goto out_pm;
+		if (ret)
+			dev_dbg(gi2c->se.dev, "GPI bus clear returned %d, issuing stop anyway\n",
+				ret);
+
+		ret = geni_i2c_gpi_bus_cmd(gi2c, I2C_STOP_ON_BUS);
+		if (ret == -ETIMEDOUT)
+			goto out_pm;
+	} else {
+		ret = geni_i2c_fifo_bus_cmd(gi2c, I2C_BUS_CLEAR);
+		if (!ret)
+			ret = geni_i2c_fifo_bus_cmd(gi2c, I2C_STOP_ON_BUS);
+	}
 
 	/*
-	 * Recovery succeeds only once both bus lines are released. The bus
-	 * state is the authority, while a clean opcode status with either
-	 * line low is a failed recovery.
+	 * Recovery succeeds when both bus lines are released. Trust the bus
+	 * state over the opcode return path, which cannot flag a benign
+	 * recovery event in GPI mode.
 	 */
 	ret = geni_i2c_bus_is_free(gi2c) ? 0 : (ret ?: -EBUSY);
 
@@ -1356,10 +1439,8 @@ static int geni_i2c_probe(struct platform_device *pdev)
 	if (ret < 0)
 		return ret;
 
-	if (!gi2c->gpi_mode) {
-		gi2c->rinfo.recover_bus = geni_i2c_recover_bus;
-		gi2c->adap.bus_recovery_info = &gi2c->rinfo;
-	}
+	gi2c->rinfo.recover_bus = geni_i2c_recover_bus;
+	gi2c->adap.bus_recovery_info = &gi2c->rinfo;
 
 	ret = i2c_add_adapter(&gi2c->adap);
 	if (ret) {

-- 
2.34.1

Re: [PATCH v4 3/3] i2c: qcom-geni: Add bus recovery support for GPI DMA mode

[Mukesh Savaliya]

On 11/3/2026 2:26 PM, Aniket Randive wrote:
> I2C transfers in GPI DMA mode can fail with -EIO when a target holds
> SDA low, leaving the bus stuck and preventing subsequent transactions.
> The GPI completion callback reports a NACK and a genuine bus error
> alike as -EIO, and the existing recovery path switches the SE to FIFO
> mode which is not valid in GPI DMA mode.
> 
> Add bus recovery support using the I2C_BUS_CLEAR and I2C_STOP_ON_BUS
> opcodes issued via the GPI TX DMA channel. Guard the -EIO recovery
> trigger with a SE_GENI_IOS I2C_BUS_FREE check to skip recovery when both
> SDA and SCL are already released. Extend the recovery callback to dispatch
> through the GPI DMA path when gi2c->gpi_mode is set.
> 
> Use the adapter timeout for GPI recovery commands, and preserve the DMA
> configuration error code when GPI setup fails.
> 
> Stop recovery on a GPI command timeout and report DMA termination
> failures. Do not issue another command or override the timeout with a
> bus-idle result, since DMA cleanup may not have completed successfully.
> 
> Signed-off-by: Aniket Randive <aniket.randive@oss.qualcomm.com>
> ---
>   drivers/i2c/busses/i2c-qcom-geni.c | 107 ++++++++++++++++++++++++++++++++-----
>   1 file changed, 94 insertions(+), 13 deletions(-)
> 
> diff --git a/drivers/i2c/busses/i2c-qcom-geni.c b/drivers/i2c/busses/i2c-qcom-geni.c

[...]

> @@ -1356,10 +1439,8 @@ static int geni_i2c_probe(struct platform_device *pdev)
>   	if (ret < 0)
>   		return ret;
>   
> -	if (!gi2c->gpi_mode) {
> -		gi2c->rinfo.recover_bus = geni_i2c_recover_bus;
> -		gi2c->adap.bus_recovery_info = &gi2c->rinfo;
> -	}
> +	gi2c->rinfo.recover_bus = geni_i2c_recover_bus;
> +	gi2c->adap.bus_recovery_info = &gi2c->rinfo;

I guess one more time this. Why patch 2 is adding and patch 3 is 
removing the change ? keep it clean way.
>   
>   	ret = i2c_add_adapter(&gi2c->adap);
>   	if (ret) {
> 

Re: [PATCH v4 3/3] i2c: qcom-geni: Add bus recovery support for GPI DMA mode

[Aniket RANDIVE]

On 11/3/2026 4:14 PM, Mukesh Savaliya wrote:
> 
> 
> On 11/3/2026 2:26 PM, Aniket Randive wrote:
>> I2C transfers in GPI DMA mode can fail with -EIO when a target holds
>> SDA low, leaving the bus stuck and preventing subsequent transactions.
>> The GPI completion callback reports a NACK and a genuine bus error
>> alike as -EIO, and the existing recovery path switches the SE to FIFO
>> mode which is not valid in GPI DMA mode.
>>
>> Add bus recovery support using the I2C_BUS_CLEAR and I2C_STOP_ON_BUS
>> opcodes issued via the GPI TX DMA channel. Guard the -EIO recovery
>> trigger with a SE_GENI_IOS I2C_BUS_FREE check to skip recovery when both
>> SDA and SCL are already released. Extend the recovery callback to 
>> dispatch
>> through the GPI DMA path when gi2c->gpi_mode is set.
>>
>> Use the adapter timeout for GPI recovery commands, and preserve the DMA
>> configuration error code when GPI setup fails.
>>
>> Stop recovery on a GPI command timeout and report DMA termination
>> failures. Do not issue another command or override the timeout with a
>> bus-idle result, since DMA cleanup may not have completed successfully.
>>
>> Signed-off-by: Aniket Randive <aniket.randive@oss.qualcomm.com>
>> ---
>>   drivers/i2c/busses/i2c-qcom-geni.c | 107 +++++++++++++++++++++++++++ 
>> +++++-----
>>   1 file changed, 94 insertions(+), 13 deletions(-)
>>
>> diff --git a/drivers/i2c/busses/i2c-qcom-geni.c b/drivers/i2c/busses/ 
>> i2c-qcom-geni.c
> 
> [...]
> 
>> @@ -1356,10 +1439,8 @@ static int geni_i2c_probe(struct 
>> platform_device *pdev)
>>       if (ret < 0)
>>           return ret;
>> -    if (!gi2c->gpi_mode) {
>> -        gi2c->rinfo.recover_bus = geni_i2c_recover_bus;
>> -        gi2c->adap.bus_recovery_info = &gi2c->rinfo;
>> -    }
>> +    gi2c->rinfo.recover_bus = geni_i2c_recover_bus;
>> +    gi2c->adap.bus_recovery_info = &gi2c->rinfo;
> 
> I guess one more time this. Why patch 2 is adding and patch 3 is 
> removing the change ? keep it clean way.

This is intentional. In patch 2, bus recovery is supported only in FIFO 
mode, so we guard the registration with !gi2c->gpi_mode. Patch 3 adds 
the corresponding bus recovery support for GPI mode, after which the 
mode check is no longer needed and is removed.

>>       ret = i2c_add_adapter(&gi2c->adap);
>>       if (ret) {
>>
> 

Re: [PATCH v4 3/3] i2c: qcom-geni: Add bus recovery support for GPI DMA mode

[Mukesh Savaliya]

On 11/3/2026 4:46 PM, Aniket RANDIVE wrote:
> 
> 
> On 11/3/2026 4:14 PM, Mukesh Savaliya wrote:
>>
>>
>> On 11/3/2026 2:26 PM, Aniket Randive wrote:
>>> I2C transfers in GPI DMA mode can fail with -EIO when a target holds
>>> SDA low, leaving the bus stuck and preventing subsequent transactions.
>>> The GPI completion callback reports a NACK and a genuine bus error
>>> alike as -EIO, and the existing recovery path switches the SE to FIFO
>>> mode which is not valid in GPI DMA mode.
>>>
>>> Add bus recovery support using the I2C_BUS_CLEAR and I2C_STOP_ON_BUS
>>> opcodes issued via the GPI TX DMA channel. Guard the -EIO recovery
>>> trigger with a SE_GENI_IOS I2C_BUS_FREE check to skip recovery when both
>>> SDA and SCL are already released. Extend the recovery callback to 
>>> dispatch
>>> through the GPI DMA path when gi2c->gpi_mode is set.
>>>
>>> Use the adapter timeout for GPI recovery commands, and preserve the DMA
>>> configuration error code when GPI setup fails.
>>>
>>> Stop recovery on a GPI command timeout and report DMA termination
>>> failures. Do not issue another command or override the timeout with a
>>> bus-idle result, since DMA cleanup may not have completed successfully.
>>>
>>> Signed-off-by: Aniket Randive <aniket.randive@oss.qualcomm.com>
>>> ---
>>>   drivers/i2c/busses/i2c-qcom-geni.c | 107 ++++++++++++++++++++++++++ 
>>> + +++++-----
>>>   1 file changed, 94 insertions(+), 13 deletions(-)
>>>
>>> diff --git a/drivers/i2c/busses/i2c-qcom-geni.c b/drivers/i2c/busses/ 
>>> i2c-qcom-geni.c
>>
>> [...]
>>
>>> @@ -1356,10 +1439,8 @@ static int geni_i2c_probe(struct 
>>> platform_device *pdev)
>>>       if (ret < 0)
>>>           return ret;
>>> -    if (!gi2c->gpi_mode) {
>>> -        gi2c->rinfo.recover_bus = geni_i2c_recover_bus;
>>> -        gi2c->adap.bus_recovery_info = &gi2c->rinfo;
>>> -    }
>>> +    gi2c->rinfo.recover_bus = geni_i2c_recover_bus;
>>> +    gi2c->adap.bus_recovery_info = &gi2c->rinfo;
>>
>> I guess one more time this. Why patch 2 is adding and patch 3 is 
>> removing the change ? keep it clean way.
> 
> This is intentional. In patch 2, bus recovery is supported only in FIFO 
> mode, so we guard the registration with !gi2c->gpi_mode. Patch 3 adds 
> the corresponding bus recovery support for GPI mode, after which the 
> mode check is no longer needed and is removed.
> 

You are right, i saw that as complete logical block removal.
>>>       ret = i2c_add_adapter(&gi2c->adap);
>>>       if (ret) {
>>>
>>
> 

Re: [PATCH v4 3/3] i2c: qcom-geni: Add bus recovery support for GPI DMA mode

[Mukesh Savaliya]

On 11/3/2026 2:26 PM, Aniket Randive wrote:
> I2C transfers in GPI DMA mode can fail with -EIO when a target holds
> SDA low, leaving the bus stuck and preventing subsequent transactions.
> The GPI completion callback reports a NACK and a genuine bus error
> alike as -EIO, and the existing recovery path switches the SE to FIFO
> mode which is not valid in GPI DMA mode.
> 
> Add bus recovery support using the I2C_BUS_CLEAR and I2C_STOP_ON_BUS
> opcodes issued via the GPI TX DMA channel. Guard the -EIO recovery
> trigger with a SE_GENI_IOS I2C_BUS_FREE check to skip recovery when both
> SDA and SCL are already released. Extend the recovery callback to dispatch
> through the GPI DMA path when gi2c->gpi_mode is set.
> 
> Use the adapter timeout for GPI recovery commands, and preserve the DMA
> configuration error code when GPI setup fails.
> 
> Stop recovery on a GPI command timeout and report DMA termination
> failures. Do not issue another command or override the timeout with a
> bus-idle result, since DMA cleanup may not have completed successfully.
> 
> Signed-off-by: Aniket Randive <aniket.randive@oss.qualcomm.com>
> ---
Reviewed-by: Mukesh Kumar Savaliya <mukesh.savaliya@oss.qualcomm.com>