dmaengine: ioat: fixing the wrong chancnt

  The chancnt would be updated in __dma_async_device_channel_register(),
but it was assigned in ioat_enumerate_channels(). Therefore chancnt has
the wrong value.

Clear chancnt before calling dma_async_device_register().

Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
---
 drivers/dma/ioat/init.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
  

On 8/11/23 01:16, Yajun Deng wrote:
> The chancnt would be updated in __dma_async_device_channel_register(),
> but it was assigned in ioat_enumerate_channels(). Therefore chancnt has
> the wrong value.
> 
> Clear chancnt before calling dma_async_device_register().
> 
> Signed-off-by: Yajun Deng <yajun.deng@linux.dev>

Thank you for the patch Yajun.

While this may work, it clobbers the chancnt read from the hardware. I 
think the preferable fix is to move the value read from the hardware in 
ioat_enumerate_channels() and its current usages to 'struct 
ioatdma_device' and leave dma->chancnt unchanged in that function so 
that zeroing it later is not needed.

Also, have you tested this patch or is this just from visual inspection?

And need a fixes tag.

> ---
>   drivers/dma/ioat/init.c | 5 ++++-
>   1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/dma/ioat/init.c b/drivers/dma/ioat/init.c
> index c4602bfc9c74..928fc8a83a36 100644
> --- a/drivers/dma/ioat/init.c
> +++ b/drivers/dma/ioat/init.c
> @@ -536,8 +536,11 @@ static int ioat_probe(struct ioatdma_device *ioat_dma)
>   
>   static int ioat_register(struct ioatdma_device *ioat_dma)
>   {
> -	int err = dma_async_device_register(&ioat_dma->dma_dev);
> +	int err;
> +
> +	ioat_dma->dma_dev.chancnt = 0;
>   
> +	err = dma_async_device_register(&ioat_dma->dma_dev);
>   	if (err) {
>   		ioat_disable_interrupts(ioat_dma);
>   		dma_pool_destroy(ioat_dma->completion_pool);
  

August 11, 2023 at 11:40 PM, "Dave Jiang" <dave.jiang@intel.com> wrote:


> 
> On 8/11/23 01:16, Yajun Deng wrote:
> 
> > 
> > The chancnt would be updated in __dma_async_device_channel_register(),
> >  but it was assigned in ioat_enumerate_channels(). Therefore chancnt has
> >  the wrong value.
> >  Clear chancnt before calling dma_async_device_register().
> >  Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
> > 
> 
> Thank you for the patch Yajun.
> 
> While this may work, it clobbers the chancnt read from the hardware. I think the preferable fix is to move the value read from the hardware in ioat_enumerate_channels() and its current usages to 'struct ioatdma_device' and leave dma->chancnt unchanged in that function so that zeroing it later is not needed.
>
Yes, it's even better. I noticed that chancnt is hardware related in ioat, so I just clear it before calling dma_async_device_register().It would be updated after calling dma_async_device_register(). And it would have
the same value with read in ioat_enumerate_channels().
It doesn't seem clobber the chancnt read from the hardware. 
 
> Also, have you tested this patch or is this just from visual inspection?
> 
Yes, I tested it.

➜  ~ ls /sys/class/dma
dma0chan0  dma1chan0  dma2chan0  dma3chan0

before:
➜  ~ cat /sys/kernel/debug/dmaengine/summary
dma0 (0000:00:04.0): number of channels: 2
dma1 (0000:00:04.1): number of channels: 2
dma2 (0000:00:04.2): number of channels: 2
dma3 (0000:00:04.3): number of channels: 2

after:
➜  ~ cat /sys/kernel/debug/dmaengine/summary
dma0 (0000:00:04.0): number of channels: 1
dma1 (0000:00:04.1): number of channels: 1
dma2 (0000:00:04.2): number of channels: 1
dma3 (0000:00:04.3): number of channels: 1


> And need a fixes tag.
>
I've tried to find the commit introduced, it looks like it was introduced from the source.
The following commits are related to chancnt：

0bbd5f4e97ff ("[I/OAT]: Driver for the Intel(R) I/OAT DMA engine")
device->common.chancnt = ioatdma_read8(device, IOAT_CHANCNT_OFFSET);

e38288117c50 ("ioatdma: Remove the wrappers around read(bwl)/write(bwl) in ioatdma")
device->common.chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);

584ec22759c0 ("ioat: move to drivers/dma/ioat/")
move driver/dma/ioatdma.c to driver/dma/ioat/

f2427e276ffe ("ioat: split ioat_dma_probe into core/version-specific routines")
dma->chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);

55f878ec47e3 ("dmaengine: ioatdma: fixup ioatdma_device namings")
dma->chancnt = readb(ioat_dma->reg_base + IOAT_CHANCNT_OFFSET);

It looks very historic. I'm confused about which one to choose. 
This is a bug, but it only affects /sys/kernel/debug/dmaengine/summary.
So I didn't add a fixes tag.

 
> > 
> > ---
> >  drivers/dma/ioat/init.c | 5 ++++-
> >  1 file changed, 4 insertions(+), 1 deletion(-)
> >  diff --git a/drivers/dma/ioat/init.c b/drivers/dma/ioat/init.c
> >  index c4602bfc9c74..928fc8a83a36 100644
> >  --- a/drivers/dma/ioat/init.c
> >  +++ b/drivers/dma/ioat/init.c
> >  @@ -536,8 +536,11 @@ static int ioat_probe(struct ioatdma_device *ioat_dma)
> >  > static int ioat_register(struct ioatdma_device *ioat_dma)
> >  {
> >  - int err = dma_async_device_register(&ioat_dma->dma_dev);
> >  + int err;
> >  +
> >  + ioat_dma->dma_dev.chancnt = 0;
> >  > + err = dma_async_device_register(&ioat_dma->dma_dev);
> >  if (err) {
> >  ioat_disable_interrupts(ioat_dma);
> >  dma_pool_destroy(ioat_dma->completion_pool);
> >
>
  

On 8/13/23 01:59, Yajun Deng wrote:
> August 11, 2023 at 11:40 PM, "Dave Jiang" <dave.jiang@intel.com> wrote:
> 
> 
>>
>> On 8/11/23 01:16, Yajun Deng wrote:
>>
>>>
>>> The chancnt would be updated in __dma_async_device_channel_register(),
>>>   but it was assigned in ioat_enumerate_channels(). Therefore chancnt has
>>>   the wrong value.
>>>   Clear chancnt before calling dma_async_device_register().
>>>   Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
>>>
>>
>> Thank you for the patch Yajun.
>>
>> While this may work, it clobbers the chancnt read from the hardware. I think the preferable fix is to move the value read from the hardware in ioat_enumerate_channels() and its current usages to 'struct ioatdma_device' and leave dma->chancnt unchanged in that function so that zeroing it later is not needed.
>>
> Yes, it's even better. I noticed that chancnt is hardware related in ioat, so I just clear it before calling dma_async_device_register().It would be updated after calling dma_async_device_register(). And it would have
> the same value with read in ioat_enumerate_channels().
> It doesn't seem clobber the chancnt read from the hardware.
>   
>> Also, have you tested this patch or is this just from visual inspection?
>>
> Yes, I tested it.
> 
> ➜  ~ ls /sys/class/dma
> dma0chan0  dma1chan0  dma2chan0  dma3chan0
> 
> before:
> ➜  ~ cat /sys/kernel/debug/dmaengine/summary
> dma0 (0000:00:04.0): number of channels: 2
> dma1 (0000:00:04.1): number of channels: 2
> dma2 (0000:00:04.2): number of channels: 2
> dma3 (0000:00:04.3): number of channels: 2
> 
> after:
> ➜  ~ cat /sys/kernel/debug/dmaengine/summary
> dma0 (0000:00:04.0): number of channels: 1
> dma1 (0000:00:04.1): number of channels: 1
> dma2 (0000:00:04.2): number of channels: 1
> dma3 (0000:00:04.3): number of channels: 1
> 
> 
>> And need a fixes tag.
>>
> I've tried to find the commit introduced, it looks like it was introduced from the source.
> The following commits are related to chancnt：
> 
> 0bbd5f4e97ff ("[I/OAT]: Driver for the Intel(R) I/OAT DMA engine")
> device->common.chancnt = ioatdma_read8(device, IOAT_CHANCNT_OFFSET);
> 
> e38288117c50 ("ioatdma: Remove the wrappers around read(bwl)/write(bwl) in ioatdma")
> device->common.chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
> 
> 584ec22759c0 ("ioat: move to drivers/dma/ioat/")
> move driver/dma/ioatdma.c to driver/dma/ioat/
> 
> f2427e276ffe ("ioat: split ioat_dma_probe into core/version-specific routines")
> dma->chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
> 
> 55f878ec47e3 ("dmaengine: ioatdma: fixup ioatdma_device namings")
> dma->chancnt = readb(ioat_dma->reg_base + IOAT_CHANCNT_OFFSET);
> 
> It looks very historic. I'm confused about which one to choose.
> This is a bug, but it only affects /sys/kernel/debug/dmaengine/summary.
> So I didn't add a fixes tag.
> 

The issue seems cosmetic and does not impact functionality. Let's just 
leave it alone then. The driver is very old.

>   
>>>
>>> ---
>>>   drivers/dma/ioat/init.c | 5 ++++-
>>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>>   diff --git a/drivers/dma/ioat/init.c b/drivers/dma/ioat/init.c
>>>   index c4602bfc9c74..928fc8a83a36 100644
>>>   --- a/drivers/dma/ioat/init.c
>>>   +++ b/drivers/dma/ioat/init.c
>>>   @@ -536,8 +536,11 @@ static int ioat_probe(struct ioatdma_device *ioat_dma)
>>>   > static int ioat_register(struct ioatdma_device *ioat_dma)
>>>   {
>>>   - int err = dma_async_device_register(&ioat_dma->dma_dev);
>>>   + int err;
>>>   +
>>>   + ioat_dma->dma_dev.chancnt = 0;
>>>   > + err = dma_async_device_register(&ioat_dma->dma_dev);
>>>   if (err) {
>>>   ioat_disable_interrupts(ioat_dma);
>>>   dma_pool_destroy(ioat_dma->completion_pool);
>>>
>>

I was thinking of something like this to keep the original functionality 
the same but let the dma->chancnt setup as intended:

diff --git a/drivers/dma/ioat/dma.h b/drivers/dma/ioat/dma.h
index 35e06b382603..a180171087a8 100644
--- a/drivers/dma/ioat/dma.h
+++ b/drivers/dma/ioat/dma.h
@@ -74,6 +74,7 @@ struct ioatdma_device {
  	struct dca_provider *dca;
  	enum ioat_irq_mode irq_mode;
  	u32 cap;
+	int chancnt;

  	/* shadow version for CB3.3 chan reset errata workaround */
  	u64 msixtba0;
diff --git a/drivers/dma/ioat/init.c b/drivers/dma/ioat/init.c
index c4602bfc9c74..9c364e92cb82 100644
--- a/drivers/dma/ioat/init.c
+++ b/drivers/dma/ioat/init.c
@@ -420,7 +420,7 @@ int ioat_dma_setup_interrupts(struct ioatdma_device 
*ioat_dma)

  msix:
  	/* The number of MSI-X vectors should equal the number of channels */
-	msixcnt = ioat_dma->dma_dev.chancnt;
+	msixcnt = ioat_dma->chancnt;
  	for (i = 0; i < msixcnt; i++)
  		ioat_dma->msix_entries[i].entry = i;

@@ -511,7 +511,7 @@ static int ioat_probe(struct ioatdma_device *ioat_dma)
  	dma_cap_set(DMA_MEMCPY, dma->cap_mask);
  	dma->dev = &pdev->dev;

-	if (!dma->chancnt) {
+	if (!ioat_dma->chancnt) {
  		dev_err(dev, "channel enumeration error\n");
  		goto err_setup_interrupts;
  	}
@@ -567,15 +567,16 @@ static void ioat_enumerate_channels(struct 
ioatdma_device *ioat_dma)
  	struct device *dev = &ioat_dma->pdev->dev;
  	struct dma_device *dma = &ioat_dma->dma_dev;
  	u8 xfercap_log;
+	int chancnt;
  	int i;

  	INIT_LIST_HEAD(&dma->channels);
-	dma->chancnt = readb(ioat_dma->reg_base + IOAT_CHANCNT_OFFSET);
-	dma->chancnt &= 0x1f; /* bits [4:0] valid */
-	if (dma->chancnt > ARRAY_SIZE(ioat_dma->idx)) {
+	chancnt = readb(ioat_dma->reg_base + IOAT_CHANCNT_OFFSET);
+	chancnt &= 0x1f; /* bits [4:0] valid */
+	if (chancnt > ARRAY_SIZE(ioat_dma->idx)) {
  		dev_warn(dev, "(%d) exceeds max supported channels (%zu)\n",
-			 dma->chancnt, ARRAY_SIZE(ioat_dma->idx));
-		dma->chancnt = ARRAY_SIZE(ioat_dma->idx);
+			 chancnt, ARRAY_SIZE(ioat_dma->idx));
+		chancnt = ARRAY_SIZE(ioat_dma->idx);
  	}
  	xfercap_log = readb(ioat_dma->reg_base + IOAT_XFERCAP_OFFSET);
  	xfercap_log &= 0x1f; /* bits [4:0] valid */
@@ -583,7 +584,7 @@ static void ioat_enumerate_channels(struct 
ioatdma_device *ioat_dma)
  		return;
  	dev_dbg(dev, "%s: xfercap = %d\n", __func__, 1 << xfercap_log);

-	for (i = 0; i < dma->chancnt; i++) {
+	for (i = 0; i < chancnt; i++) {
  		ioat_chan = kzalloc(sizeof(*ioat_chan), GFP_KERNEL);
  		if (!ioat_chan)
  			break;
@@ -596,7 +597,7 @@ static void ioat_enumerate_channels(struct 
ioatdma_device *ioat_dma)
  			break;
  		}
  	}
-	dma->chancnt = i;
+	ioat_dma->chancnt = i;
  }

  /**