[PATCH v2] mtd: rawnand: denali: add more delays before latching incoming data

[PATCH v2] mtd: rawnand: denali: add more delays before latching incoming data

[Masahiro Yamada]

The Denali IP have several registers to specify how many clock cycles
should be waited between falling/rising signals. You can improve the
NAND access performance by programming these registers with optimized
values.

Because struct nand_sdr_timings represents the device requirement
in pico seconds, denali_setup_data_interface() computes the register
values by dividing the device timings with the clock period.

Marek Vasut reported this driver in the latest kernel does not work
on his SOCFPGA board. (The on-board NAND chip is mode 5)

The suspicious parameter is acc_clks, so this commit relaxes it.

The Denali NAND Flash Memory Controller User's Guide describes this
register as follows:

  acc_clks
    signifies the number of bus interface clk_x clock cycles,
    controller should wait from read enable going low to sending
    out a strobe of clk_x for capturing of incoming data.

Currently, acc_clks is calculated only based on tREA, the delay on the
chip side. This does not include additional delays that come from the
data path on the PCB and in the SoC, load capacity of the pins, etc.

This relatively becomes a big factor on faster timing modes like mode 5.

Before supporting the ->setup_data_interface() hook (e.g. Linux 4.12),
the Denali driver hacks acc_clks in a couple of ways [1] [2] to support
the timing mode 5.

We would not go back to the hard-coded acc_clks, but we need to include
this factor into the delay somehow. Let's say the amount of the additional
delay is 10000 pico sec.

In the new calculation, acc_clks is determined by timings->tREA_max +
data_setup_on_host.

Also, prolong the RE# low period to make sure the data hold is met.

Finally, re-center the data latch timing for extra safety.

[1] https://github.com/torvalds/linux/blob/v4.12/drivers/mtd/nand/denali.c#L276
[2] https://github.com/torvalds/linux/blob/v4.12/drivers/mtd/nand/denali.c#L282

Reported-by: Marek Vasut <marex@denx.de>
Signed-off-by: Masahiro Yamada <yamada.masahiro@socionext.com>
---

Changes in v2:
  - Use 'unsigned int' instead of 'u32' for data_setup_on_host
  - Fix 'line over 80 characters' warning from checkpatch.pl

 drivers/mtd/nand/raw/denali.c | 45 ++++++++++++++++++++++++++---------
 1 file changed, 34 insertions(+), 11 deletions(-)

diff --git a/drivers/mtd/nand/raw/denali.c b/drivers/mtd/nand/raw/denali.c
index 6a6c919b2569..2fcd2baf6e35 100644
--- a/drivers/mtd/nand/raw/denali.c
+++ b/drivers/mtd/nand/raw/denali.c
@@ -764,6 +764,7 @@ static int denali_write_page(struct nand_chip *chip, const u8 *buf,
 static int denali_setup_data_interface(struct nand_chip *chip, int chipnr,
 				       const struct nand_data_interface *conf)
 {
+	static const unsigned int data_setup_on_host = 10000;
 	struct denali_controller *denali = to_denali_controller(chip);
 	struct denali_chip_sel *sel;
 	const struct nand_sdr_timings *timings;
@@ -796,15 +797,6 @@ static int denali_setup_data_interface(struct nand_chip *chip, int chipnr,
 
 	sel = &to_denali_chip(chip)->sels[chipnr];
 
-	/* tREA -> ACC_CLKS */
-	acc_clks = DIV_ROUND_UP(timings->tREA_max, t_x);
-	acc_clks = min_t(int, acc_clks, ACC_CLKS__VALUE);
-
-	tmp = ioread32(denali->reg + ACC_CLKS);
-	tmp &= ~ACC_CLKS__VALUE;
-	tmp |= FIELD_PREP(ACC_CLKS__VALUE, acc_clks);
-	sel->acc_clks = tmp;
-
 	/* tRWH -> RE_2_WE */
 	re_2_we = DIV_ROUND_UP(timings->tRHW_min, t_x);
 	re_2_we = min_t(int, re_2_we, RE_2_WE__VALUE);
@@ -862,14 +854,45 @@ static int denali_setup_data_interface(struct nand_chip *chip, int chipnr,
 	tmp |= FIELD_PREP(RDWR_EN_HI_CNT__VALUE, rdwr_en_hi);
 	sel->rdwr_en_hi_cnt = tmp;
 
-	/* tRP, tWP -> RDWR_EN_LO_CNT */
+	/*
+	 * tREA -> ACC_CLKS
+	 * tRP, tWP, tRHOH, tRC, tWC -> RDWR_EN_LO_CNT
+	 */
+
+	/*
+	 * Determine the minimum of acc_clks to meet the setup timing when
+	 * capturing the incoming data.
+	 *
+	 * The delay on the chip side is well-defined as tREA, but we need to
+	 * take additional delay into account. This includes a certain degree
+	 * of unknowledge, such as signal propagation delays on the PCB and
+	 * in the SoC, load capacity of the I/O pins, etc.
+	 */
+	acc_clks = DIV_ROUND_UP(timings->tREA_max + data_setup_on_host, t_x);
+
+	/* Determine the minimum of rdwr_en_lo_cnt from RE#/WE# pulse width */
 	rdwr_en_lo = DIV_ROUND_UP(max(timings->tRP_min, timings->tWP_min), t_x);
+
+	/* Extend rdwr_en_lo to meet the data hold timing */
+	rdwr_en_lo = max_t(int, rdwr_en_lo,
+			   acc_clks - timings->tRHOH_min / t_x);
+
+	/* Extend rdwr_en_lo to meet the requirement for RE#/WE# cycle time */
 	rdwr_en_lo_hi = DIV_ROUND_UP(max(timings->tRC_min, timings->tWC_min),
 				     t_x);
-	rdwr_en_lo_hi = max_t(int, rdwr_en_lo_hi, mult_x);
 	rdwr_en_lo = max(rdwr_en_lo, rdwr_en_lo_hi - rdwr_en_hi);
 	rdwr_en_lo = min_t(int, rdwr_en_lo, RDWR_EN_LO_CNT__VALUE);
 
+	/* Center the data latch timing for extra safety */
+	acc_clks = (acc_clks + rdwr_en_lo +
+		    DIV_ROUND_UP(timings->tRHOH_min, t_x)) / 2;
+	acc_clks = min_t(int, acc_clks, ACC_CLKS__VALUE);
+
+	tmp = ioread32(denali->reg + ACC_CLKS);
+	tmp &= ~ACC_CLKS__VALUE;
+	tmp |= FIELD_PREP(ACC_CLKS__VALUE, acc_clks);
+	sel->acc_clks = tmp;
+
 	tmp = ioread32(denali->reg + RDWR_EN_LO_CNT);
 	tmp &= ~RDWR_EN_LO_CNT__VALUE;
 	tmp |= FIELD_PREP(RDWR_EN_LO_CNT__VALUE, rdwr_en_lo);
-- 
2.17.1

Re: [PATCH v2] mtd: rawnand: denali: add more delays before latching incoming data

[Marek Vasut]

On 3/17/20 8:18 AM, Masahiro Yamada wrote:
> The Denali IP have several registers to specify how many clock cycles
> should be waited between falling/rising signals. You can improve the
> NAND access performance by programming these registers with optimized
> values.
> 
> Because struct nand_sdr_timings represents the device requirement
> in pico seconds, denali_setup_data_interface() computes the register
> values by dividing the device timings with the clock period.
> 
> Marek Vasut reported this driver in the latest kernel does not work
> on his SOCFPGA board. (The on-board NAND chip is mode 5)
> 
> The suspicious parameter is acc_clks, so this commit relaxes it.
> 
> The Denali NAND Flash Memory Controller User's Guide describes this
> register as follows:
> 
>   acc_clks
>     signifies the number of bus interface clk_x clock cycles,
>     controller should wait from read enable going low to sending
>     out a strobe of clk_x for capturing of incoming data.
> 
> Currently, acc_clks is calculated only based on tREA, the delay on the
> chip side. This does not include additional delays that come from the
> data path on the PCB and in the SoC, load capacity of the pins, etc.
> 
> This relatively becomes a big factor on faster timing modes like mode 5.
> 
> Before supporting the ->setup_data_interface() hook (e.g. Linux 4.12),
> the Denali driver hacks acc_clks in a couple of ways [1] [2] to support
> the timing mode 5.
> 
> We would not go back to the hard-coded acc_clks, but we need to include
> this factor into the delay somehow. Let's say the amount of the additional
> delay is 10000 pico sec.
> 
> In the new calculation, acc_clks is determined by timings->tREA_max +
> data_setup_on_host.
> 
> Also, prolong the RE# low period to make sure the data hold is met.
> 
> Finally, re-center the data latch timing for extra safety.
> 
> [1] https://github.com/torvalds/linux/blob/v4.12/drivers/mtd/nand/denali.c#L276
> [2] https://github.com/torvalds/linux/blob/v4.12/drivers/mtd/nand/denali.c#L282
> 
> Reported-by: Marek Vasut <marex@denx.de>
> Signed-off-by: Masahiro Yamada <yamada.masahiro@socionext.com>

I tested it on the AV SoCFPGA, this seems to work, so feel free to apply.

Re: [PATCH v2] mtd: rawnand: denali: add more delays before latching incoming data

[Miquel Raynal]

Hi Marek,

Marek Vasut <marex@denx.de> wrote on Wed, 22 Apr 2020 17:29:53 +0200:

> On 3/17/20 8:18 AM, Masahiro Yamada wrote:
> > The Denali IP have several registers to specify how many clock cycles
> > should be waited between falling/rising signals. You can improve the
> > NAND access performance by programming these registers with optimized
> > values.
> > 
> > Because struct nand_sdr_timings represents the device requirement
> > in pico seconds, denali_setup_data_interface() computes the register
> > values by dividing the device timings with the clock period.
> > 
> > Marek Vasut reported this driver in the latest kernel does not work
> > on his SOCFPGA board. (The on-board NAND chip is mode 5)
> > 
> > The suspicious parameter is acc_clks, so this commit relaxes it.
> > 
> > The Denali NAND Flash Memory Controller User's Guide describes this
> > register as follows:
> > 
> >   acc_clks
> >     signifies the number of bus interface clk_x clock cycles,
> >     controller should wait from read enable going low to sending
> >     out a strobe of clk_x for capturing of incoming data.
> > 
> > Currently, acc_clks is calculated only based on tREA, the delay on the
> > chip side. This does not include additional delays that come from the
> > data path on the PCB and in the SoC, load capacity of the pins, etc.
> > 
> > This relatively becomes a big factor on faster timing modes like mode 5.
> > 
> > Before supporting the ->setup_data_interface() hook (e.g. Linux 4.12),
> > the Denali driver hacks acc_clks in a couple of ways [1] [2] to support
> > the timing mode 5.
> > 
> > We would not go back to the hard-coded acc_clks, but we need to include
> > this factor into the delay somehow. Let's say the amount of the additional
> > delay is 10000 pico sec.
> > 
> > In the new calculation, acc_clks is determined by timings->tREA_max +
> > data_setup_on_host.
> > 
> > Also, prolong the RE# low period to make sure the data hold is met.
> > 
> > Finally, re-center the data latch timing for extra safety.
> > 
> > [1] https://github.com/torvalds/linux/blob/v4.12/drivers/mtd/nand/denali.c#L276
> > [2] https://github.com/torvalds/linux/blob/v4.12/drivers/mtd/nand/denali.c#L282
> > 
> > Reported-by: Marek Vasut <marex@denx.de>
> > Signed-off-by: Masahiro Yamada <yamada.masahiro@socionext.com>  
> 
> I tested it on the AV SoCFPGA, this seems to work, so feel free to apply.


Great! Thanks a lot for testing, would you mind sending your Tested-by?

Thanks,
Miquèl

Re: [PATCH v2] mtd: rawnand: denali: add more delays before latching incoming data

[Marek Vasut]

On 4/22/20 5:36 PM, Miquel Raynal wrote:
> Hi Marek,
> 
> Marek Vasut <marex@denx.de> wrote on Wed, 22 Apr 2020 17:29:53 +0200:
> 
>> On 3/17/20 8:18 AM, Masahiro Yamada wrote:
>>> The Denali IP have several registers to specify how many clock cycles
>>> should be waited between falling/rising signals. You can improve the
>>> NAND access performance by programming these registers with optimized
>>> values.
>>>
>>> Because struct nand_sdr_timings represents the device requirement
>>> in pico seconds, denali_setup_data_interface() computes the register
>>> values by dividing the device timings with the clock period.
>>>
>>> Marek Vasut reported this driver in the latest kernel does not work
>>> on his SOCFPGA board. (The on-board NAND chip is mode 5)
>>>
>>> The suspicious parameter is acc_clks, so this commit relaxes it.
>>>
>>> The Denali NAND Flash Memory Controller User's Guide describes this
>>> register as follows:
>>>
>>>   acc_clks
>>>     signifies the number of bus interface clk_x clock cycles,
>>>     controller should wait from read enable going low to sending
>>>     out a strobe of clk_x for capturing of incoming data.
>>>
>>> Currently, acc_clks is calculated only based on tREA, the delay on the
>>> chip side. This does not include additional delays that come from the
>>> data path on the PCB and in the SoC, load capacity of the pins, etc.
>>>
>>> This relatively becomes a big factor on faster timing modes like mode 5.
>>>
>>> Before supporting the ->setup_data_interface() hook (e.g. Linux 4.12),
>>> the Denali driver hacks acc_clks in a couple of ways [1] [2] to support
>>> the timing mode 5.
>>>
>>> We would not go back to the hard-coded acc_clks, but we need to include
>>> this factor into the delay somehow. Let's say the amount of the additional
>>> delay is 10000 pico sec.
>>>
>>> In the new calculation, acc_clks is determined by timings->tREA_max +
>>> data_setup_on_host.
>>>
>>> Also, prolong the RE# low period to make sure the data hold is met.
>>>
>>> Finally, re-center the data latch timing for extra safety.
>>>
>>> [1] https://github.com/torvalds/linux/blob/v4.12/drivers/mtd/nand/denali.c#L276
>>> [2] https://github.com/torvalds/linux/blob/v4.12/drivers/mtd/nand/denali.c#L282
>>>
>>> Reported-by: Marek Vasut <marex@denx.de>
>>> Signed-off-by: Masahiro Yamada <yamada.masahiro@socionext.com>  
>>
>> I tested it on the AV SoCFPGA, this seems to work, so feel free to apply.
> 
> 
> Great! Thanks a lot for testing, would you mind sending your Tested-by?

Lightly
Tested-by: Marek Vasut <marex@denx.de>

Re: [PATCH v2] mtd: rawnand: denali: add more delays before latching incoming data

[Miquel Raynal]

On Tue, 2020-03-17 at 07:18:21 UTC, Masahiro Yamada wrote:
> The Denali IP have several registers to specify how many clock cycles
> should be waited between falling/rising signals. You can improve the
> NAND access performance by programming these registers with optimized
> values.
> 
> Because struct nand_sdr_timings represents the device requirement
> in pico seconds, denali_setup_data_interface() computes the register
> values by dividing the device timings with the clock period.
> 
> Marek Vasut reported this driver in the latest kernel does not work
> on his SOCFPGA board. (The on-board NAND chip is mode 5)
> 
> The suspicious parameter is acc_clks, so this commit relaxes it.
> 
> The Denali NAND Flash Memory Controller User's Guide describes this
> register as follows:
> 
>   acc_clks
>     signifies the number of bus interface clk_x clock cycles,
>     controller should wait from read enable going low to sending
>     out a strobe of clk_x for capturing of incoming data.
> 
> Currently, acc_clks is calculated only based on tREA, the delay on the
> chip side. This does not include additional delays that come from the
> data path on the PCB and in the SoC, load capacity of the pins, etc.
> 
> This relatively becomes a big factor on faster timing modes like mode 5.
> 
> Before supporting the ->setup_data_interface() hook (e.g. Linux 4.12),
> the Denali driver hacks acc_clks in a couple of ways [1] [2] to support
> the timing mode 5.
> 
> We would not go back to the hard-coded acc_clks, but we need to include
> this factor into the delay somehow. Let's say the amount of the additional
> delay is 10000 pico sec.
> 
> In the new calculation, acc_clks is determined by timings->tREA_max +
> data_setup_on_host.
> 
> Also, prolong the RE# low period to make sure the data hold is met.
> 
> Finally, re-center the data latch timing for extra safety.
> 
> [1] https://github.com/torvalds/linux/blob/v4.12/drivers/mtd/nand/denali.c#L276
> [2] https://github.com/torvalds/linux/blob/v4.12/drivers/mtd/nand/denali.c#L282
> 
> Reported-by: Marek Vasut <marex@denx.de>
> Signed-off-by: Masahiro Yamada <yamada.masahiro@socionext.com>
> Tested-by: Marek Vasut <marex@denx.de>

Applied to https://git.kernel.org/pub/scm/linux/kernel/git/mtd/linux.git nand/next, thanks.

Miquel

Re: [PATCH v2] mtd: rawnand: denali: add more delays before latching incoming data

[Ralph Siemsen]

On Tue, Mar 17, 2020 at 04:18:21PM +0900, Masahiro Yamada wrote:
>
>Marek Vasut reported this driver in the latest kernel does not work
>on his SOCFPGA board. (The on-board NAND chip is mode 5)

In a bit of an ironic twist, this change seems to cause a regression for 
me. I'm also using a Cyclone V SoC system, similar but not the same as 
the SOCFPGA eval board. The NAND device in my case is S34ML04G2.

It worked fine in 4.9, 4.19 and 5.4 kernels. However after upgrading to 
5.10 the NAND driver fails with "timeout while waiting for irq 0x4", 
just as was reported back in 2018:
https://lore.kernel.org/linux-mtd/737ejrj81y.fsf@pengutronix.de/

Reverting commit 5756f2e8dad46eba6e2d3e530243b8eff4dd5a42 restores the 
behaviour in my case, but that would presumably break the SOCFPGA board.

Any thoughts on how we could make it work for everybody?

Ralph

Re: [PATCH v2] mtd: rawnand: denali: add more delays before latching incoming data

[Masahiro Yamada]

On Thu, Jul 1, 2021 at 3:47 AM Ralph Siemsen <ralph.siemsen@linaro.org> wrote:
>
> On Tue, Mar 17, 2020 at 04:18:21PM +0900, Masahiro Yamada wrote:
> >
> >Marek Vasut reported this driver in the latest kernel does not work
> >on his SOCFPGA board. (The on-board NAND chip is mode 5)
>
> In a bit of an ironic twist, this change seems to cause a regression for
> me. I'm also using a Cyclone V SoC system, similar but not the same as
> the SOCFPGA eval board. The NAND device in my case is S34ML04G2.
>
> It worked fine in 4.9, 4.19 and 5.4 kernels. However after upgrading to
> 5.10 the NAND driver fails with "timeout while waiting for irq 0x4",
> just as was reported back in 2018:
> https://lore.kernel.org/linux-mtd/737ejrj81y.fsf@pengutronix.de/
>
> Reverting commit 5756f2e8dad46eba6e2d3e530243b8eff4dd5a42 restores the
> behaviour in my case, but that would presumably break the SOCFPGA board.
>
> Any thoughts on how we could make it work for everybody?
>
> Ralph


I no longer have any hardware to test this driver.

If you increase data_setup_on_host (the current value is 10000),
will it solve your issue?





-- 
Best Regards
Masahiro Yamada

Re: [PATCH v2] mtd: rawnand: denali: add more delays before latching incoming data

[Ralph Siemsen]

On Mon, Jul 05, 2021 at 01:00:43AM +0900, Masahiro Yamada wrote:
>
>I no longer have any hardware to test this driver.
>
>If you increase data_setup_on_host (the current value is 10000),
>will it solve your issue?

Yes, it does seem to work with this value increased to 10001.

Looking more closely, I compared timing parameters in struct 
denali_chip_sel, for the kernel 5.4 (which works), for 5.10 (which 
fails), and also for 5.10 with data_setup_host=10001 (which works).

                 5.4     5.10    5.10+10001
acc_clks        2       3       4
rdwr_en_lo_cnt  3       2       3
cs_setup_cnt    1       0       0

So it seems that on my hardware, rdwr_en_lo_cnt must be >= 3.

Regards,
Ralph

Re: [PATCH v2] mtd: rawnand: denali: add more delays before latching incoming data

[Helen]

On Mon, Jul 05, 2021 at 14:48 PM Ralph Siemsen <ralph.siemsen@linaro.org> wrote:
>
>Yes, it does seem to work with this value increased to 10001.
>
>Looking more closely, I compared timing parameters in struct 
>denali_chip_sel, for the kernel 5.4 (which works), for 5.10 (which 
>fails), and also for 5.10 with data_setup_host=10001 (which works).

>                 5.4     5.10    5.10+10001
>acc_clks        2       3       4
>rdwr_en_lo_cnt  3       2       3
>cs_setup_cnt    1       0       0
>
>So it seems that on my hardware, rdwr_en_lo_cnt must be >= 3.

I'm using a Cyclone V SoC system, the NAND device is Micron MT29F8G08ABACAWP,
NAND driver fails with "timeout while waiting for irq 0x4", which found in 2021:
https://lore.kernel.org/linux-mtd/20210705144823.GA1459782@maple.netwinder.org/

It worked fine in Mode5 in 5.10 kernels without the commit below:
https://github.com/torvalds/linux/commit/5756f2e8dad46eba6e2d3e530243b8eff4dd5a42
But after upgrading to 6.12 kernel, include this commit with data_setup_on_host=10001,
the NAND driver still fails with "timeout while waiting for irq 0x4".

Add some print to timing calculations, it seems the deletion of the line in the commit:
-	rdwr_en_lo_hi = max_t(int, rdwr_en_lo_hi, mult_x);
cause rdwr_en_lo_hi changed from 4 to 1 and indirectly changed rdwr_en_lo from 3 to 2.
It seems on my hardware rdwr_en_lo_cnt must be >= 3, nand works if add this line back.
I'd like to know why delete it?

Linux 5.10 without commit:
rdwr_en_hi = DIV_ROUD_UP(max(timings->tREH_min, timings->tWH_min), t_x);
       0x1 = (7000 + 20000 - 1)/20000 = 
acc_clks = DIV_ROUND_UP(timings->tREA_max, t_x);
     0x1 = DIV_ROUND_UP(16000, 20000);
rdwr_en_lo = DIV_ROUND_UP(max(timings->tRP_min, timings->tWP_min), t_x);
       0x1 = DIV_ROUND_UP(max(10000, 10000), 20000);
rdwr_en_lo_hi = DIV_ROUND_UP(max(timings->tRC_min, timings->tWC_min), t_x);
          0x1 = DIV_ROUND_UP(max(2000, 2000), 20000);
rdwr_en_lo_hi = max_t(int, rdwr_en_lo_hi, mult_x);
          0x4 = max_t(int, 0x1, 0x4);
rdwr_en_lo = max(rdwr_en_lo, rdwr_en_lo_hi - rdwr_en_hi);
       0x3 = max(0x1, 0x4 - 0x1)
rdwr_en_lo = min_t(int, rdwr_en_lo, RDWR_EN_LO_CNT__VALUE);
       0x3 = min_t(int, 0x3, mask);

Linux 6.12 with commit + 10001:
static const unsigned int data_setup_on_host = 10001;
rdwr_en_hi = DIV_ROUD_UP(max(timings->tREH_min, timings->tWH_min), t_x);
       0x1 = (7000 + 20000 - 1)/20000
acc_clks = DIV_ROUND_UP(timings->tREA_max + data_setup_on_host, t_x);
     0x2 = (16000 + 10001 + 20000 - 1)/20000
rdwr_en_lo = DIV_ROUND_UP(max(timings->tRP_min, timings->tWP_min), t_x);
       0x1 = (10000 + 20000 -1)/20000
rdwr_en_lo = max_t(int, rdwr_en_lo, acc_clks - timings->tRHOH_min / t_x);
       0x2 = max_t(int, 0x1, 0x2 - 15000 / 20000);
rdwr_en_lo_hi = DIV_ROUND_UP(max(timings->tRC_min, timings->tWC_min), t_x);
          0x1 = (20000 + 20000 - 1)/20000
rdwr_en_lo = max(rdwr_en_lo, rdwr_en_lo_hi - rdwr_en_hi);
       0x2 = max(0x2, 0x2 - 0x1);
rdwr_en_lo = min_t(int, rdwr_en_lo, RDWR_EN_LO_CNT__VALUE);
       0x2 = min_t(int, 0x2, mask);

BR,
Helen