Re: [PATCH 2/2] ARM: dts: imx7: add DMA properties for ECSPI

[Stefan Agner]

[adding Mark Brown]

On 11.02.2019 02:23, Shawn Guo wrote:
> On Thu, Feb 07, 2019 at 09:00:44PM +0000, Trent Piepho wrote:
>> On Mon, 2019-01-07 at 14:22 +0100, Stefan Agner wrote:
>> > Allow to use DMA for SPI by adding the appropriate DMA properites
>> > to the ecspi nodes.
>> >
>> > Signed-off-by: Stefan Agner <stefan@xxxxxxxx>
>> > ---
>> >  arch/arm/boot/dts/imx7s.dtsi | 8 ++++++++
>> >  1 file changed, 8 insertions(+)
>> >
>> > diff --git a/arch/arm/boot/dts/imx7s.dtsi b/arch/arm/boot/dts/imx7s.dtsi
>> > index a052198f6e96..b9330176c3af 100644
>> > --- a/arch/arm/boot/dts/imx7s.dtsi
>> > +++ b/arch/arm/boot/dts/imx7s.dtsi
>> > @@ -653,6 +653,8 @@
>> >  				clocks = <&clks IMX7D_ECSPI4_ROOT_CLK>,
>> >  					<&clks IMX7D_ECSPI4_ROOT_CLK>;
>> >  				clock-names = "ipg", "per";
>> > +				dmas = <&sdma 6 7 1>, <&sdma 7 7 2>;
>> > +				dma-names = "rx", "tx";
>> >  				status = "disabled";
>> >  			};
>>
>> After updating my kernel to linux-next on my IMX7d based device, I
>> found that an FPGA, which is programmed via the ECSPI interface, was no
>> longer accepting its image.
>>
>> I tracked the problem to this change.  If I turn off DMA, it works.
>>
>> There's an interesting thing that happens when DMA is used.  The SPI
>> clock changes.  Instead of cycling continuously for the entire
>> transfer, it instead clocks out 8 bits, then pauses for 4 bit times,
>> then the next byte, etc.  So it's a net of about 50% slower.  The pause
>> between bytes scales with spi frequency to always be about 4 bits.
>>
>> Here's a trace with DMA: https://imagebin.ca/v/4WEkEnvsVSkq
>>
>> Here's what it looks like without DMA:
>> https://imagebin.ca/v/4WEkVfEqpQ12
>>
>> It seems like there are other problems with DMA too.  Here's an error
>> I'll random get every so often.
>>
>> [  142.082325] spi_master spi1: I/O Error in DMA RX
>> [  142.085678] spidev spi1.0: SPI transfer failed: -110
>> [  142.089389] spi_master spi1: failed to transfer one message from queue
>>
>> Not sure if the timeout is overly aggressive or if there is some other failure.
>>
>> Then sometimes there errors are worse:
>>
>> Internal error: Oops - undefined instruction: 0 [#1] PREEMPT SMP ARM
>> Modules linked in:
>> CPU: 0 PID: 1006 Comm: fpga-loader Not tainted 5.0.0-rc4-next-20190201 #1
>> Hardware name: Freescale i.MX7 Dual (Device Tree)
>> PC is at sg_last+0x4c/0x68
>> LR is at   (null)
>> pc : [<80494800>]    lr : [<00000000>]    psr: 60010013
>> sp : bf0add5c  ip : bea94598  fp : bf371218
>> r10: bf0949f0  r9 : 00000000  r8 : bf094800
>> r7 : bdc39300  r6 : bf371000  r5 : bdc39300  r4 : bf094b68
>> r3 : 00000000  r2 : 00000001  r1 : 00000001  r0 : bea94580
>> Flags: nZCv  IRQs on  FIQs on  Mode SVC_32  ISA ARM  Segment user
>> Control: 30c5387d  Table: bea94000  DAC: fffffffd
>> Process fpga-loader (pid: 1006, stack limit = 0x61bc8b73)
>> Stack: (0xbf0add5c to 0xbf0ae000)
>> dd40:                                                                8056cdb0
>> dd60: 8056c9ac 00001000 00000000 bdc39300 bf094800 bf094b68 bf371000 00000000
>> dd80: bf0949f0 8056bc00 bf094800 bdc39300 bf0adeac bf094800 bf094ad0 80569edc
>> ...
>> dfc0: 00217301 00150000 00001000 00000036 00df2010 00000003 00001000 00df1008
>> dfe0: 00022f70 7e9ebb14 000113b8 76f2a7f8 20010030 00000003 00000000 00000000
>> [<80494800>] (sg_last) from [<8056cdb0>] (spi_imx_transfer+0xd8/0x448)
>> [<8056cdb0>] (spi_imx_transfer) from [<8056bc00>] (spi_bitbang_transfer_one+0x50/0xa0)
>> [<8056bc00>] (spi_bitbang_transfer_one) from [<80569edc>] (spi_transfer_one_message+0x18c/0x3e0)
>> [<80569edc>] (spi_transfer_one_message) from [<8056a498>] (__spi_pump_messages+0x368/0x518)
>> [<8056a498>] (__spi_pump_messages) from [<8056a7ec>] (__spi_sync+0x198/0x1a0)
>> [<8056a7ec>] (__spi_sync) from [<8056a818>] (spi_sync+0x24/0x3c)
>> [<8056a818>] (spi_sync) from [<8056ae14>] (spidev_sync+0x38/0x4c)
>> [<8056ae14>] (spidev_sync) from [<8056b6f4>] (spidev_ioctl+0x660/0x704)
>> [<8056b6f4>] (spidev_ioctl) from [<80335f9c>] (do_vfs_ioctl+0xac/0x79c)
>> [<80335f9c>] (do_vfs_ioctl) from [<803366c0>] (ksys_ioctl+0x34/0x58)
>> [<803366c0>] (ksys_ioctl) from [<80201120>] (ret_fast_syscall+0x0/0x4c)
>> Exception stack(0xbf0adfa8 to 0xbf0adff0)
>> dfa0:                   00217301 00150000 00000003 40206b00 7e9ebb80 03938700
>> dfc0: 00217301 00150000 00001000 00000036 00df2010 00000003 00001000 00df1008
>> dfe0: 00022f70 7e9ebb14 000113b8 76f2a7f8
>> Code: e1510002 1afffff3 e3530000 149df004 (e7f001f2)
>> ---[ end trace 1588229fc7541669 ]---
>>
>> I think DMA on imx not be ready for prime time yet.

I did only a few transfers at 10MHz on an i.MX 6ULL and i.MX 7 device
those seemed to work. I did load the RAM firmware.

The i.MX 6 device trees have DMA since quite some time, so it seems to
work fine there? (maybe Sascha can confirm, he fixed device tree DMA
properties in imx6qdl.dtsi a while ago).


> 
> I dropped both patches from my tree.
> 

I think this is the wrong approach to disable DMA on those devices. The
device tree is supposed to describe the complete hardware. If the driver
is not ready to support DMA for that particular variant, we should add
this information in the driver. We have compatible strings for i.MX
6UL/i.MX 7 to disable DMA accordingly.

--
Stefan