Linux-mtd Archive on
 help / color / Atom feed
From: "Cédric Le Goater" <>
To: Boris Brezillon <>
Cc: Mark Rutland <>,
	Vignesh Raghavendra <>,
	linux-aspeed <>,
	Tudor Ambarus <>,
	Andrew Jeffery <>,
	Richard Weinberger <>,
	Marek Vasut <>,, Joel Stanley <>,
	Miquel Raynal <>,
	Brian Norris <>,
	David Woodhouse <>,
	Linux ARM <>
Subject: Re: [PATCH 00/16] mtd: spi-nor: aspeed: AST2600 support and extensions
Date: Fri, 11 Oct 2019 15:07:18 +0200
Message-ID: <> (raw)
In-Reply-To: <>

On 11/10/2019 14:07, Boris Brezillon wrote:
> On Fri, 11 Oct 2019 13:47:53 +0200
> Cédric Le Goater <> wrote:
>> On 11/10/2019 11:51, Boris Brezillon wrote:
>>> On Fri, 11 Oct 2019 11:29:49 +0200
>>> Cédric Le Goater <> wrote:
>>>> On 11/10/2019 08:45, Boris Brezillon wrote:  
>>>>> On Thu, 10 Oct 2019 23:47:45 +0000
>>>>> Joel Stanley <> wrote:
>>>>>> On Wed, 9 Oct 2019 at 20:56, Boris Brezillon
>>>>>> <> wrote:    
>>>>>>> Hi Cedric,
>>>>>>> On Fri,  4 Oct 2019 13:59:03 +0200
>>>>>>> Cédric Le Goater <> wrote:
>>>>>>>> Hello,
>>>>>>>> This series first extends the support for the Aspeed AST2500 and
>>>>>>>> AST2400 SMC driver. It adds Dual Data support and read training giving
>>>>>>>> the best read settings for a given chip. Support for the new AST2600
>>>>>>>> SoC is added at the end.
>>>>>>>> I understand that a new spi_mem framework exists and I do have an
>>>>>>>> experimental driver using it. But unfortunately, it is difficult to
>>>>>>>> integrate the read training. The Aspeed constraints are not compatible
>>>>>>>> and i haven't had the time to extend the current framework.      
>>>>>>> Hm, I don't think that's a good reason to push new features to the
>>>>>>> existing driver, especially since I asked others to migrate their
>>>>>>> drivers to spi-mem in the past. I do understand your concerns, and I'll
>>>>>>> let the SPI NOR/MTD maintainers make the final call, but I think it'd
>>>>>>> be better for the SPI MEM ecosystem to think about this link-training
>>>>>>> API (Vignesh needs it for the Cadence driver IIRC) rather than pushing
>>>>>>> this kind of feature to spi-nor controller drivers.      
>>>>>> As Cedric mentioned, the OpenBMC project has been shipping the read
>>>>>> training code for the ast2400/ast2400 for several years now. It would
>>>>>> be great to see it in mainline.
>>>>>> I think it's reasonable to ask for the driver to be moved to the
>>>>>> spi-mem subsystem once it has the required APIs.    
>>>>> Except it won't have the necessary APIs unless someone works on it, and
>>>>> adding this feature to existing spi-nor drivers won't help achieving
>>>>> this goal.    
>>>> What would you suggest ? Something like the patch below which would
>>>> call a 'train' operation at the end of spi_add_device().  
>>> This has been discussed in the past with Vignesh, but I can't find the
>>> thread where this discussion happened. My understanding was that link
>>> training would use a command with well-known output (is there a
>>> dedicated SPI NOR command for that?) and test different clk settings
>>> until it finds one that works.  
>> The read training on Aspeed consists of finding the appropriate read
>> timing delays for well-known HCLK dividers and store the results in 
>> the Read Timing Compensation register. 
>> We first read a golden buffer at low speed and then perform reads with 
>> different clocks and delay cycle settings to find a breaking point.
> Where does this golden buffer come from? Do you have a specific command
> to access it? Is it a NOR section with well-known data?

We read the first 16K of the flash device at low speed.  

>> This 
>> selects the default clock frequency for the CEx control register. 
>>>> Also, when doing read training, we might need to know some lowlevel 
>>>> characteristics of the chip being trained. Should we offer a way 
>>>> to grab the probed m25p80 device and give access to the underlying 
>>>> 'struct spi_nor' ? 
>>>>   static struct spi_nor *spi_get_pnor(struct spi_device *spi)
>>>>   {
>>>> 	struct spi_mem *spimem = spi_get_drvdata(spi);
>>>> 	struct m25p *flash = spi_mem_get_drvdata(spimem);
>>>> 	return flash ? &flash->spi_nor : NULL;
>>>>   }
>>>> Yeah, it's hideous. I just want to raise the issue.  
>>> Oh no. We definitely don't want to expose the spi_nor chip to the
>>> spi_mem layer, but, if needed, we can add more fields to spi_mem and
>>> let the spi_mem driver fill them. We just need to figure out what's
>>> really needed.  
>> We need the SPI/NOR flash characteristics for link training and its 
>> size to configure the controller to access the CS on the AHB window. 
> We managed to deal with direct mapping without having to explicitly pass
> the NOR size (we just pass the size of the direct mapping we want to
> create). What do you need the size for? 

because the AHB window on which are mapped all CS is segmented. 

There is one sub window for each CS and all segments need to be 
configured in the controller (in the segment registers). In case 
of a bad value or an overlap, you can loose access to the CS or 
hang the system.

To access the SPI NOR registers, you don't need a window of the 
exact size because in that case the controller operates in 'User' 
command mode and the window only needs to be partially opened. 
Reads and Writes use the start address of AHB sub-window of the CS.

If you want access to the full contents through a direct mapping,
the controller operates in 'Read' or 'Fast Read' command mode and
the window needs to be at least as wide as the flash size.

> Is it just to configure the AHB window for the link-training stuff? 

The link-training uses the direct mapping mode but not on the
whole contents, only on the first 16K. So the AHB window does not 
have to be as wide as the chip and we could use a default setting. 
But generally, we already have good knowledge on the chip before
training is started.

> If that's the case, I guess this can be part of the op_template 
> I was talking about, or maybe passed as extra parameters.
>> [ ... ]
>>>>  	int (*fw_translate_cs)(struct spi_controller *ctlr, unsigned cs);
>>>> +
>>>> +	int			(*train)(struct spi_device *spi);  
>>> Was more thinking of something like:
>>> 	int (*link_setup)(struct spi_mem *mem,
>>> 			  struct spi_mem_op *op_template,
>>> 			  ...);
>>> where the op_template would potentially differ depending on the type of
>>> memory (NOR, NAND, SRAM?). I also don't know what other params would be
>>> needed to do the link training.  
>> The Aspeed controller needs to set read delay timings at different HCLK
>> settings. It's doing read operations with the default IO mode.
> So you need information about the theoretical read delay so you can
> adjust them, right. I guess that's something we can pass to the
> ->link_setup() hook.

The read delay settings really depend on the controller. I was thinking 
more of a post_setup() hook called once in m25p_probe() after the chip 
is scanned. 

As for the parameter, we would need some SPI-NOR info or some generic
SPI device data that the controller would know how to interpret.



>>> BTW, this hook should be in the spi_mem_controller_ops struct not in
>>> spi_controller.  
>> ok. Let's wait for feedback from Vinesh.
> Thanks for starting this discussion.
> Boris

Linux MTD discussion mailing list

  reply index

Thread overview: 34+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-10-04 11:59 Cédric Le Goater
2019-10-04 11:59 ` [PATCH 01/16] mtd: spi-nor: aspeed: Use command mode for reads Cédric Le Goater
2019-10-04 11:59 ` [PATCH 02/16] mtd: spi-nor: aspeed: Add support for SPI dual IO read mode Cédric Le Goater
2019-10-04 11:59 ` [PATCH 03/16] mtd: spi-nor: aspeed: Link controller with the ahb clock Cédric Le Goater
2019-10-04 11:59 ` [PATCH 04/16] mtd: spi-nor: aspeed: Add read training Cédric Le Goater
2019-10-11 12:28   ` Boris Brezillon
2019-10-11 13:13     ` Vignesh Raghavendra
2019-10-11 14:03       ` Cédric Le Goater
2019-10-11 13:55     ` Cédric Le Goater
2019-10-11 14:29       ` Boris Brezillon
2019-10-11 14:37         ` Cédric Le Goater
2019-10-04 11:59 ` [PATCH 05/16] mtd: spi-nor: aspeed: Limit the maximum SPI frequency Cédric Le Goater
2019-10-04 11:59 ` [PATCH 06/16] mtd: spi-nor: fix options for mx66l51235f Cédric Le Goater
2019-10-04 16:23   ` Cédric Le Goater
2019-10-04 11:59 ` [PATCH 07/16] mtd: spi-nor: aspeed: Add support for the 4B opcodes Cédric Le Goater
2019-10-04 11:59 ` [PATCH 08/16] mtd: spi-nor: Add support for w25q512jv Cédric Le Goater
2019-10-04 11:59 ` [PATCH 09/16] mtd: spi-nor: aspeed: Introduce a field for the AHB physical address Cédric Le Goater
2019-10-04 11:59 ` [PATCH 10/16] mtd: spi-nor: aspeed: Introduce segment operations Cédric Le Goater
2019-10-04 11:59 ` [PATCH 11/16] dt-bindings: mtd: aspeed-smc: Add new comptatible for AST2600 Cédric Le Goater
2019-10-15 19:26   ` Rob Herring
2019-10-04 11:59 ` [PATCH 12/16] mtd: spi-nor: aspeed: Add initial support for the AST2600 Cédric Le Goater
2019-10-04 11:59 ` [PATCH 13/16] mtd: spi-nor: aspeed: Check for disabled segments on " Cédric Le Goater
2019-10-04 12:09 ` [PATCH 14/16] mtd: spi-nor: aspeed: Introduce training operations per platform Cédric Le Goater
2019-10-04 12:09   ` [PATCH 15/16] mtd: spi-nor: aspeed: Introduce a HCLK mask for training Cédric Le Goater
2019-10-04 12:09   ` [PATCH 16/16] mtd: spi-nor: aspeed: Add read training support for the AST2600 Cédric Le Goater
2019-10-09 20:55 ` [PATCH 00/16] mtd: spi-nor: aspeed: AST2600 support and extensions Boris Brezillon
2019-10-10 23:47   ` Joel Stanley
2019-10-11  6:45     ` Boris Brezillon
2019-10-11  9:29       ` Cédric Le Goater
2019-10-11  9:51         ` Boris Brezillon
2019-10-11 11:47           ` Cédric Le Goater
2019-10-11 12:07             ` Boris Brezillon
2019-10-11 13:07               ` Cédric Le Goater [this message]
2019-10-11 14:01                 ` Boris Brezillon

Reply instructions:

You may reply publically to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

Linux-mtd Archive on

Archives are clonable:
	git clone --mirror linux-mtd/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 linux-mtd linux-mtd/ \
	public-inbox-index linux-mtd

Example config snippet for mirrors

Newsgroup available over NNTP:

AGPL code for this site: git clone