Re: [PATCH v4 2/3] dma: arm-dma350: support combined IRQ mode with runtime IRQ topology detection
From: Robin Murphy <robin.murphy@arm.com>
Date: 2026-03-24 12:59:31
Also in:
dmaengine, linux-devicetree, lkml
On 2026-03-23 11:48 am, Jun Guo wrote:
quoted hunk ↗ jump to hunk
DMA-350 can be integrated with either per-channel IRQ lines or a single combined IRQ line. Add support for both layouts in a unified way. Detect IRQ topology at probe time via platform_irq_count(), then: - request one global IRQ and enable DMANSECCTRL.INTREN_ANYCHINTR for combined mode, or - request per-channel IRQs for channel mode. Refactor IRQ completion/error handling into a shared channel handler used by both global and per-channel IRQ paths, and guard against IRQs arriving without an active descriptor. Assisted-by: Cursor: GPT-5.3-Codex Signed-off-by: Jun Guo <redacted> --- drivers/dma/arm-dma350.c | 165 +++++++++++++++++++++++++++++++++------ 1 file changed, 139 insertions(+), 26 deletions(-)diff --git a/drivers/dma/arm-dma350.c b/drivers/dma/arm-dma350.c index 84220fa83029..2cf6f783b44f 100644 --- a/drivers/dma/arm-dma350.c +++ b/drivers/dma/arm-dma350.c@@ -14,6 +14,7 @@ #include "virt-dma.h" #define DMAINFO 0x0f00 +#define DRIVER_NAME "arm-dma350"
Why is this here in the middle of register definitions? But frankly why is it being added at all?
quoted hunk ↗ jump to hunk
#define DMA_BUILDCFG0 0xb0 #define DMA_CFG_DATA_WIDTH GENMASK(18, 16)@@ -142,6 +143,9 @@ #define LINK_LINKADDR BIT(30) #define LINK_LINKADDRHI BIT(31) +/* DMA NONSECURE CONTROL REGISTER */ +#define DMANSECCTRL 0x20c
Please follow the existing code rather than making a confusing ugly mess. DMANSECCTRL is a register frame at offset 0x0200; it contains (among others) a register named NSEC_CTRL at offset 0x0c.
quoted hunk ↗ jump to hunk
+#define INTREN_ANYCHINTR_EN BIT(0) enum ch_ctrl_donetype { CH_CTRL_DONETYPE_NONE = 0,@@ -192,6 +196,7 @@ struct d350_chan { struct d350 { struct dma_device dma; + void __iomem *base;
Why?
quoted hunk ↗ jump to hunk
int nchan; int nreq; struct d350_chan channels[] __counted_by(nchan);@@ -461,18 +466,40 @@ static void d350_issue_pending(struct dma_chan *chan) spin_unlock_irqrestore(&dch->vc.lock, flags); } -static irqreturn_t d350_irq(int irq, void *data) +static void d350_handle_chan_irq(struct d350_chan *dch, struct device *dev, + int chan_id, u32 ch_status) { - struct d350_chan *dch = data; - struct device *dev = dch->vc.chan.device->dev; - struct virt_dma_desc *vd = &dch->desc->vd; - u32 ch_status; + struct virt_dma_desc *vd; + bool intr_done = ch_status & CH_STAT_INTR_DONE; + bool intr_err = ch_status & CH_STAT_INTR_ERR; - ch_status = readl(dch->base + CH_STATUS); - if (!ch_status) - return IRQ_NONE; + if (!intr_done && !intr_err) {
Why is the logic being changed here?
+ if (chan_id >= 0) + dev_warn(dev, "Channel %d unexpected IRQ: 0x%08x\n", + chan_id, ch_status);
Why? I see no reason why "dev" shouldn't still be the channel device.
+ else
+ dev_warn(dev, "Unexpected IRQ source? 0x%08x\n", ch_status);
+ writel_relaxed(ch_status, dch->base + CH_STATUS);
+ return;
+ }
+
+ writel_relaxed(ch_status, dch->base + CH_STATUS);
+
+ spin_lock(&dch->vc.lock);
+ if (!dch->desc) {
+ if (chan_id >= 0)
+ dev_warn(dev,
+ "Channel %d IRQ without active descriptor: 0x%08x\n",
+ chan_id, ch_status);
+ else
+ dev_warn(dev, "IRQ without active descriptor: 0x%08x\n",
+ ch_status);
+ spin_unlock(&dch->vc.lock);
+ return;
+ }This seems entirely new and nothing to do with refactoring. Why do you think it's necessary?
quoted hunk ↗ jump to hunk
- if (ch_status & CH_STAT_INTR_ERR) { + vd = &dch->desc->vd; + if (intr_err) { u32 errinfo = readl_relaxed(dch->base + CH_ERRINFO); if (errinfo & (CH_ERRINFO_AXIRDPOISERR | CH_ERRINFO_AXIRDRESPERR))@@ -483,14 +510,10 @@ static irqreturn_t d350_irq(int irq, void *data) vd->tx_result.result = DMA_TRANS_ABORTED; vd->tx_result.residue = d350_get_residue(dch); - } else if (!(ch_status & CH_STAT_INTR_DONE)) { - dev_warn(dev, "Unexpected IRQ source? 0x%08x\n", ch_status); } - writel_relaxed(ch_status, dch->base + CH_STATUS); - spin_lock(&dch->vc.lock);
I can't recall the details off-hand right now, but I have a strong feeling the order of operations here was reasonably significant - what's the justifcation for changing it?
quoted hunk ↗ jump to hunk
vchan_cookie_complete(vd); - if (ch_status & CH_STAT_INTR_DONE) { + if (intr_done) { dch->status = DMA_COMPLETE; dch->residue = 0; d350_start_next(dch);@@ -499,6 +522,44 @@ static irqreturn_t d350_irq(int irq, void *data) dch->residue = vd->tx_result.residue; } spin_unlock(&dch->vc.lock); +} + +static irqreturn_t d350_global_irq(int irq, void *data) +{ + struct d350 *dmac = (struct d350 *)data;
This isn't C++
+ irqreturn_t ret = IRQ_NONE; + int i; + + (void)irq;
Why?
+ for (i = 0; i < dmac->nchan; i++) {
+ struct d350_chan *dch = &dmac->channels[i];
+ u32 ch_status;
+
+ ch_status = readl(dch->base + CH_STATUS);
+ if (!ch_status)Why is this factored out, only to be duplicated in both callers?
quoted hunk ↗ jump to hunk
+ continue; + + ret = IRQ_HANDLED; + d350_handle_chan_irq(dch, dmac->dma.dev, i, ch_status); + } + + return ret; +} + +static irqreturn_t d350_channel_irq(int irq, void *data) +{ + struct d350_chan *dch = data; + struct device *dev = dch->vc.chan.device->dev; + u32 ch_status; + + (void)irq; + + ch_status = readl(dch->base + CH_STATUS); + if (!ch_status) + return IRQ_NONE; + + d350_handle_chan_irq(dch, dev, -1, ch_status); return IRQ_HANDLED; }@@ -506,10 +567,18 @@ static irqreturn_t d350_irq(int irq, void *data) static int d350_alloc_chan_resources(struct dma_chan *chan) { struct d350_chan *dch = to_d350_chan(chan); - int ret = request_irq(dch->irq, d350_irq, IRQF_SHARED, - dev_name(&dch->vc.chan.dev->device), dch); - if (!ret) - writel_relaxed(CH_INTREN_DONE | CH_INTREN_ERR, dch->base + CH_INTREN); + int ret = 0; + + if (dch->irq >= 0) { + ret = request_irq(dch->irq, d350_channel_irq, IRQF_SHARED, + dev_name(&dch->vc.chan.dev->device), dch); + if (ret) { + dev_err(chan->device->dev, "Failed to request IRQ %d\n", dch->irq); + return ret; + } + } + + writel_relaxed(CH_INTREN_DONE | CH_INTREN_ERR, dch->base + CH_INTREN); return ret; }@@ -519,18 +588,21 @@ static void d350_free_chan_resources(struct dma_chan *chan) struct d350_chan *dch = to_d350_chan(chan); writel_relaxed(0, dch->base + CH_INTREN); - free_irq(dch->irq, dch); + if (dch->irq >= 0) { + free_irq(dch->irq, dch); + dch->irq = -EINVAL; + } vchan_free_chan_resources(&dch->vc); } static int d350_probe(struct platform_device *pdev) { struct device *dev = &pdev->dev; - struct d350 *dmac; + struct d350 *dmac = NULL;
Why?
quoted hunk ↗ jump to hunk
void __iomem *base; u32 reg; - int ret, nchan, dw, aw, r, p; - bool coherent, memset; + int ret, nchan, dw, aw, r, p, irq_count; + bool coherent, memset, combined_irq; base = devm_platform_ioremap_resource(pdev, 0); if (IS_ERR(base))@@ -556,6 +628,7 @@ static int d350_probe(struct platform_device *pdev) return -ENOMEM; dmac->nchan = nchan; + dmac->base = base; reg = readl_relaxed(base + DMAINFO + DMA_BUILDCFG1); dmac->nreq = FIELD_GET(DMA_CFG_NUM_TRIGGER_IN, reg);@@ -582,12 +655,46 @@ static int d350_probe(struct platform_device *pdev) dmac->dma.device_issue_pending = d350_issue_pending; INIT_LIST_HEAD(&dmac->dma.channels); + irq_count = platform_irq_count(pdev); + if (irq_count < 0) + return dev_err_probe(dev, irq_count, + "Failed to count interrupts\n"); + + if (irq_count == 1) { + combined_irq = true;
What if there's only 1 channel and it is actually using its dedicated IRQ?
+ } else if (irq_count >= nchan) {
+ combined_irq = false;
+ } else {
+ return dev_err_probe(dev, -EINVAL,
+ "Invalid IRQ count %d for %d channels\n",
+ irq_count, nchan);Thsi is a change in behaviour, since we wouldm't currently consider it an error if the last 1 or more channels don't have an interrupt, but aren't accessible anyway - if a system *always* reserved the last couple of channels for the Secure world, arguably there would be no real requiremnt to describe them in the Non-Secure DT at all.
quoted hunk ↗ jump to hunk
+ } + + if (combined_irq) { + int host_irq = platform_get_irq(pdev, 0); + + if (host_irq < 0) + return dev_err_probe(dev, host_irq, + "Failed to get IRQ\n"); + + ret = devm_request_irq(&pdev->dev, host_irq, d350_global_irq, + IRQF_SHARED, DRIVER_NAME, dmac); + if (ret) + return dev_err_probe( + dev, ret, + "Failed to request the combined IRQ %d\n", + host_irq); + /* Combined Non-Secure Channel Interrupt Enable */ + writel_relaxed(INTREN_ANYCHINTR_EN, dmac->base + DMANSECCTRL); + } + /* Would be nice to have per-channel caps for this... */ memset = true; for (int i = 0; i < nchan; i++) { struct d350_chan *dch = &dmac->channels[i]; dch->base = base + DMACH(i); + dch->irq = -EINVAL; writel_relaxed(CH_CMD_CLEAR, dch->base + CH_CMD); reg = readl_relaxed(dch->base + CH_BUILDCFG1);@@ -595,10 +702,15 @@ static int d350_probe(struct platform_device *pdev) dev_warn(dev, "No command link support on channel %d\n", i); continue; } - dch->irq = platform_get_irq(pdev, i); - if (dch->irq < 0) - return dev_err_probe(dev, dch->irq, - "Failed to get IRQ for channel %d\n", i); + + if (!combined_irq) { + dch->irq = platform_get_irq(pdev, i);
But here's the thing - even if we *were* to do something special in the DT binding, why not simply have: if (combined_irq) dch->irq = combined_irq_num; else dch->irq = platform_get_irq(pdev, i); at this point, and then let everything else keep working the way it already does? Thanks, Robin.
quoted hunk ↗ jump to hunk
+ if (dch->irq < 0) + return dev_err_probe( + dev, dch->irq, + "Failed to get IRQ for channel %d\n", + i); + } dch->has_wrap = FIELD_GET(CH_CFG_HAS_WRAP, reg); dch->has_trig = FIELD_GET(CH_CFG_HAS_TRIGIN, reg) &@@ -640,6 +752,7 @@ static void d350_remove(struct platform_device *pdev) } static const struct of_device_id d350_of_match[] __maybe_unused = { + { .compatible = "cix,sky1-dma-350" }, { .compatible = "arm,dma-350" }, {} };