Re: [PATCH v2 06/12] mtd: rawnand: stm32_fmc2: use FMC2_TIMEOUT_MS for timeouts

From: Christophe Kerello
Date: Wed Apr 29 2020 - 06:14:35 EST




On 4/29/20 12:06 PM, Miquel Raynal wrote:
Hi Christophe,

Christophe Kerello <christophe.kerello@xxxxxx> wrote on Wed, 29 Apr
2020 11:41:44 +0200:

On 4/29/20 11:35 AM, Miquel Raynal wrote:
Hi Christophe,

Christophe Kerello <christophe.kerello@xxxxxx> wrote on Wed, 29 Apr
2020 11:27:43 +0200:
Hi MiquÃl,

On 4/27/20 8:22 PM, Miquel Raynal wrote:
Hi Christophe,

Christophe Kerello <christophe.kerello@xxxxxx> wrote on Wed, 15 Apr
2020 17:57:30 +0200:
>>>> This patch removes the constant FMC2_TIMEOUT_US.
FMC2_TIMEOUT_MS is set to 5 seconds and this constant is used
each time that we need to wait (except when the timeout value
is set by the framework)

Signed-off-by: Christophe Kerello <christophe.kerello@xxxxxx>
---
drivers/mtd/nand/raw/stm32_fmc2_nand.c | 11 +++++------
1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/drivers/mtd/nand/raw/stm32_fmc2_nand.c b/drivers/mtd/nand/raw/stm32_fmc2_nand.c
index ab53314..f159c39 100644
--- a/drivers/mtd/nand/raw/stm32_fmc2_nand.c
+++ b/drivers/mtd/nand/raw/stm32_fmc2_nand.c
@@ -37,8 +37,7 @@
/* Max ECC buffer length */
#define FMC2_MAX_ECC_BUF_LEN (FMC2_BCHDSRS_LEN * FMC2_MAX_SG)
>> -#define FMC2_TIMEOUT_US 1000
-#define FMC2_TIMEOUT_MS 1000
+#define FMC2_TIMEOUT_MS 5000
>> /* Timings */
#define FMC2_THIZ 1
@@ -525,9 +524,9 @@ static int stm32_fmc2_ham_calculate(struct nand_chip *chip, const u8 *data,
u32 sr, heccr;
int ret;
>> - ret = readl_relaxed_poll_timeout(fmc2->io_base + FMC2_SR,
- sr, sr & FMC2_SR_NWRF, 10,
- FMC2_TIMEOUT_MS);
+ ret = readl_relaxed_poll_timeout_atomic(fmc2->io_base + FMC2_SR,
+ sr, sr & FMC2_SR_NWRF, 1,
+ 1000 * FMC2_TIMEOUT_MS);

Is the _atomic suffix needed here? If yes it would deserve a separate
patch with Fixes/Stable tags.
>>
I have currently not seen any issues. So, I will remove this modification as we will move to regmap_read_poll_timeout in patch 10.
if (ret) {
dev_err(fmc2->dev, "ham timeout\n");
return ret;
@@ -1315,7 +1314,7 @@ static int stm32_fmc2_waitrdy(struct nand_chip *chip, unsigned long timeout_ms)
/* Check if there is no pending requests to the NAND flash */
if (readl_relaxed_poll_timeout_atomic(fmc2->io_base + FMC2_SR, sr,
sr & FMC2_SR_NWRF, 1,
- FMC2_TIMEOUT_US))
+ 1000 * FMC2_TIMEOUT_MS))
dev_warn(fmc2->dev, "Waitrdy timeout\n");
>> /* Wait tWB before R/B# signal is low */

You change the timeouts from 1ms to 5s.

Maybe 5s is a little bit too much IMHO but we don't really care as this
is a timeout. However 1ms is tight. If you are changing this value
because it triggers error (eg. when the machine is loaded), then it is
a fix and should appear like it.

Thanks,
MiquÃl
>>
No errors currently happens.
During our stress tests, in a overloaded system, we have seen that we could be close to 1 second, even if we never met this value.
So, to be safe, I have set this timeout to 5 seconds.
As it is just a timeout value, I have not seen any side effect.
I am using the same timeout constant to avoid to have one timeout per cases.

Something is wrong in my mind:
You say you observe delays of almost up to 1 second, but the polling
currently happens on 1000 us = 1ms, either you had timeouts or I
misread something?

Thanks,
MiquÃl

Hi MiquÃl,

My fault. For this polling, we never met 1 ms.
The 1 second observed was on the sequencer when we read/write a page (as it the same timeout value that is used)

OK I get it. So perhaps you can give these details in the commit log to
explain why you use 5 seconds instead of one.

Thanks,
MiquÃl


Hi MiquÃl,

A proposal could also be to split this patch:
- a first patch that is using only one timeout value.
- a second patch that is increasing the value to 5 seconds.

Regards,
Christophe Kerello.