Skip to contentSkip to conversation

dw_dmac: don't wait for FIFO_EMPTY endlessly in dwc_chan_pause

PATCH In review Andy Shevchenko 3 participants 3 messages Mar 21, 2013 dmaengine

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 dw_dmac: don't wait for FIFO_EMPTY endlessly in dwc_chan_pause

When we pause the channel after transfer is completed we might stuck in the dwc_chan_pause() because the FIFO_EMPTY flag will never be asserted. To avoid the endless loop we introduce a timeout here (*). The proper solution is to somehow get the residue in FIFO and avoid busyloop when transfer is done, but this task is not simple and fast.

Unfortunately we can't use cpu_relax() in conjunction with jiffies checker, due to we have interrupts disabled by spin_lock_irqsave() and there is a big chance that no interrupts will come to update the jiffies..

(*) The worst case is
AHB write * FIFO size / hclk = 5.12 us,
where
AHB write = 2 cycles,
hclk = 100 MHz,
burst size = 1 byte,
FIFO size = 256 bytes.
The proposed 40us timeout might be considered as a big one, though we enter to that state only when we have the transfer already completed.

---
 drivers/dma/dw_dmac.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/dma/dw_dmac.c +3 -2
index 43a5329..43e2e89 100644
@@ -1030,10 +1030,11 @@ set_runtime_config(struct dma_chan *chan, struct dma_slave_config *sconfig)
 static inline void dwc_chan_pause(struct dw_dma_chan *dwc)
 {
 	u32 cfglo = channel_readl(dwc, CFG_LO);
+	unsigned int count = 20;	/* timeout iterations */
 
 	channel_writel(dwc, CFG_LO, cfglo | DWC_CFGL_CH_SUSP);
-	while (!(channel_readl(dwc, CFG_LO) & DWC_CFGL_FIFO_EMPTY))
-		cpu_relax();
+	while (!(channel_readl(dwc, CFG_LO) & DWC_CFGL_FIFO_EMPTY) && count--)
+		udelay(2);
 
 	dwc->paused = true;
 }
Signature · 1 line
1.8.2.rc0.22.gb3600c3
Acked-by on v1 1/1
#1 lore raw
Quoted 30 lines · Andy Shevchenko

When we pause the channel after transfer is completed we might stuck in the dwc_chan_pause() because the FIFO_EMPTY flag will never be asserted. To avoid the endless loop we introduce a timeout here (*). The proper solution is to somehow get the residue in FIFO and avoid busyloop when transfer is done, but this task is not simple and fast.

Unfortunately we can't use cpu_relax() in conjunction with jiffies checker, due to we have interrupts disabled by spin_lock_irqsave() and there is a big chance that no interrupts will come to update the jiffies..

(*) The worst case is
AHB write * FIFO size / hclk = 5.12 us,
where
AHB write = 2 cycles,
hclk = 100 MHz,
burst size = 1 byte,
FIFO size = 256 bytes.
The proposed 40us timeout might be considered as a big one, though we enter to that state only when we have the transfer already completed.

---
 drivers/dma/dw_dmac.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/dma/dw_dmac.c +0 -0
index 43a5329..43e2e89 100644

drivers/dma/dw_dmac.c set_runtime_config(struct dma_chan *chan, struct dma_slave_config *sconfig) · line 1037

Quoted 10 lines
drivers/dma/dw_dmac.c +3 -2
@@ -1030,10 +1030,11 @@ set_runtime_config(struct dma_chan *chan, struct dma_slave_config *sconfig)
 static inline void dwc_chan_pause(struct dw_dma_chan *dwc)
 {
        u32 cfglo = channel_readl(dwc, CFG_LO);
+       unsigned int count = 20;        /* timeout iterations */
 
        channel_writel(dwc, CFG_LO, cfglo | DWC_CFGL_CH_SUSP);
-       while (!(channel_readl(dwc, CFG_LO) & DWC_CFGL_FIFO_EMPTY))
-               cpu_relax();
+       while (!(channel_readl(dwc, CFG_LO) & DWC_CFGL_FIFO_EMPTY) && count--)
+               udelay(2);
#2 lore raw
In reply to Andy Shevchenko
Quoted 20 lines · Andy Shevchenko

When we pause the channel after transfer is completed we might stuck in the dwc_chan_pause() because the FIFO_EMPTY flag will never be asserted. To avoid the endless loop we introduce a timeout here (*). The proper solution is to somehow get the residue in FIFO and avoid busyloop when transfer is done, but this task is not simple and fast.

Unfortunately we can't use cpu_relax() in conjunction with jiffies checker, due to we have interrupts disabled by spin_lock_irqsave() and there is a big chance that no interrupts will come to update the jiffies..

(*) The worst case is
AHB write * FIFO size / hclk = 5.12 us,
where
AHB write = 2 cycles,
hclk = 100 MHz,
burst size = 1 byte,
FIFO size = 256 bytes.
The proposed 40us timeout might be considered as a big one, though we enter to that state only when we have the transfer already completed.

Quoted 1 line

Applied Thanks

Quoted 25 lines
---
 drivers/dma/dw_dmac.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/dma/dw_dmac.c +3 -2
index 43a5329..43e2e89 100644
@@ -1030,10 +1030,11 @@ set_runtime_config(struct dma_chan *chan, struct dma_slave_config *sconfig)
 static inline void dwc_chan_pause(struct dw_dma_chan *dwc)
 {
 	u32 cfglo = channel_readl(dwc, CFG_LO);
+	unsigned int count = 20;	/* timeout iterations */
 
 	channel_writel(dwc, CFG_LO, cfglo | DWC_CFGL_CH_SUSP);
-	while (!(channel_readl(dwc, CFG_LO) & DWC_CFGL_FIFO_EMPTY))
-		cpu_relax();
+	while (!(channel_readl(dwc, CFG_LO) & DWC_CFGL_FIFO_EMPTY) && count--)
+		udelay(2);
 
 	dwc->paused = true;
 }
Signature · 1 line
1.8.2.rc0.22.gb3600c3