Re: [PATCH v5 3/5] mtd: Add support for HyperBus memory devices

[Vignesh Raghavendra]

On 18/06/19 1:15 AM, Sergei Shtylyov wrote:
> Hello!
> 
> On 06/11/2019 02:57 PM, Vignesh Raghavendra wrote:
> 
>>>> Cypress' HyperBus is Low Signal Count, High Performance Double Data Rate
>>>> Bus interface between a host system master and one or more slave
>>>> interfaces. HyperBus is used to connect microprocessor, microcontroller,
>>>> or ASIC devices with random access NOR flash memory (called HyperFlash)
>>>> or self refresh DRAM (called HyperRAM).
>>>>
>>>> Its a 8-bit data bus (DQ[7:0]) with  Read-Write Data Strobe (RWDS)
>>>> signal and either Single-ended clock(3.0V parts) or Differential clock
>>>> (1.8V parts). It uses ChipSelect lines to select b/w multiple slaves.
>>>> At bus level, it follows a separate protocol described in HyperBus
>>>> specification[1].
>>>>
>>>> HyperFlash follows CFI AMD/Fujitsu Extended Command Set (0x0002) similar
>>>> to that of existing parallel NORs. Since HyperBus is x8 DDR bus,
>>>> its equivalent to x16 parallel NOR flash wrt bits per clock cycle. But
>>>> HyperBus operates at >166MHz frequencies.
>>>> HyperRAM provides direct random read/write access to flash memory
>>>> array.
>>>>
>>>> But, HyperBus memory controllers seem to abstract implementation details
>>>> and expose a simple MMIO interface to access connected flash.
>>>>
>>>> Add support for registering HyperFlash devices with MTD framework. MTD
>>>> maps framework along with CFI chip support framework are used to support
>>>> communicating with flash.
>>>>
>>>> Framework is modelled along the lines of spi-nor framework. HyperBus
>>>> memory controller (HBMC) drivers calls hyperbus_register_device() to
>>>> register a single HyperFlash device. HyperFlash core parses MMIO access
>>>> information from DT, sets up the map_info struct, probes CFI flash and
>>>> registers it with MTD framework.
>>>>
>>>> Some HBMC masters need calibration/training sequence[3] to be carried
>>>> out, in order for DLL inside the controller to lock, by reading a known
>>>> string/pattern. This is done by repeatedly reading CFI Query
>>>> Identification String. Calibration needs to be done before trying to detect
>>>> flash as part of CFI flash probe.
>>>>
>>>> HyperRAM is not supported at the moment.
>>>>
>>>> HyperBus specification can be found at[1]
>>>> HyperFlash datasheet can be found at[2]
>>>>
>>>> [1] https://www.cypress.com/file/213356/download
>>>> [2] https://www.cypress.com/file/213346/download
>>>> [3] http://www.ti.com/lit/ug/spruid7b/spruid7b.pdf
>>>>     Table 12-5741. HyperFlash Access Sequence
>>>>
>>>> Signed-off-by: Vignesh Raghavendra <vigneshr@xxxxxx>
>>> [...]
>>>> diff --git a/drivers/mtd/hyperbus/hyperbus-core.c b/drivers/mtd/hyperbus/hyperbus-core.c
>>>> new file mode 100644
>>>> index 000000000000..df1f75e10b1a
>>>> --- /dev/null
>>>> +++ b/drivers/mtd/hyperbus/hyperbus-core.c
>>>> @@ -0,0 +1,191 @@
>>>> +// SPDX-License-Identifier: GPL-2.0
>>>> +//
>>>> +// Copyright (C) 2019 Texas Instruments Incorporated - http://www.ti.com/
>>>> +// Author: Vignesh Raghavendra <vigneshr@xxxxxx>
>>>> +
>>>> +#include <linux/err.h>
>>>> +#include <linux/kernel.h>
>>>> +#include <linux/module.h>
>>>> +#include <linux/mtd/hyperbus.h>
>>>> +#include <linux/mtd/map.h>
>>>> +#include <linux/mtd/mtd.h>
>>>> +#include <linux/mtd/cfi.h>
>>>> +#include <linux/of.h>
>>>> +#include <linux/of_address.h>
>>>> +#include <linux/types.h>
>>>> +
>>>> +#define HYPERBUS_CALIB_COUNT 25
>>>
>>>    Mhm, I think I've already protested about this being #define'd here...
>>
>> I thought you had agreed that default optional calibration routine can
>> be part of core code and thus this #define.
>>
>> Anyways, what is your preference here? Drop the constant and use a local
>> variable in hyperbus_calibrate()?
>> Or are you suggesting to move hyperbus_calibrate() TI's specific driver?
> 
>    I'm just not comfortable with the common HF code using quite an arbitrary
> constant...
> 

Ok, I will move the code over to TI driver. We can always bring it back
to core code if more drivers need it.

>>> [...]
>>>> diff --git a/include/linux/mtd/hyperbus.h b/include/linux/mtd/hyperbus.h
>>>> new file mode 100644
>>>> index 000000000000..ee2eefd822c9
>>>> --- /dev/null
>>>> +++ b/include/linux/mtd/hyperbus.h
>>>> @@ -0,0 +1,91 @@
> [...]
>>>> + * @mtd: pointer to MTD struct
>>>> + * @ctlr: pointer to HyperBus controller struct
>>>> + * @memtype: type of memory device: HyperFlash or HyperRAM
>>>> + * @registered: flag to indicate whether device is registered with MTD core
>>>> + */
>>>> +
>>>> +struct hyperbus_device {
>>>> +	struct map_info map;
>>>> +	struct device_node *np;
>>>> +	struct mtd_info *mtd;
>>>> +	struct hyperbus_ctlr *ctlr;
>>>> +	enum hyperbus_memtype memtype;
>>>> +	bool registered;
>>>> +};
>>>> +
>>>> +/**
>>>> + * struct hyperbus_ops - struct representing custom HyperBus operations
>>>> + * @read16: read 16 bit of data, usually from register/ID-CFI space
>>>> + * @write16: write 16 bit of data, usually to register/ID-CFI space
>>>
>>>    Usually? How to differ the register/memory transfers if both are possible?
> 
>> CFI + map framework does not provide a way to differentiate b/w reg
>> access vs memory access. read16()/write16() is used to either access
>> registers or for sending various cmds like lock/unlock etc or for
>> programming a single word.
>> For regular read/writes copy_from() and copy_to() are used.
> 
>    In my case only copy_from() would exist -- no proper acceleration for
> writes...
> 

Actually copy_to() is not used by cfi_cmdset_0002.c, its always
write16() that used to program flash.
This is something I want to extend support to, so as to use DMA for
writes as well because I see that writes seem extremely slow at least on
my platform.

>> Looking at HyperBus protocol, controllers would not need to
>> differentiate b/w registers vs memory transfers for HyperFlash devices.
>> So, I think I can drop read16/write16 and redirect these calls to
>> copy_from()/copy_to()
> 
>    Doubt it, frankly speaking.

Sorry for confusion, as I said above, we do need to keep write16().
Also, copy_to maps to memcpy_toio in case of simple_map which may not
use 16 bit IO accessors. So write16() cannot be mapped to copy_to()

So we need at least write16(), copy_from() and mostly copy_to() (for
accelerating writes). So, lets keep this simple and have all map ops
including read16() as is.


> 
>> I mainly added these functions keeping HyperRAM in mind. Idea was
>> drivers would look at hyperbus_device->memtype and set to register
>> access mode for HyperRAM in case of write16()/read16(). Looks like the
>> interface is not intuitive enough
>> So, will drop these and add it back when adding HyperRAM support.
>>
>> Does that work for your HW as well?
> 
>    Don't think so...
> 
>    However, my HyperFlash driver could make use of the following #define's in
> the HyperBus header:
> 
> #define HF_CMD_CA47		BIT(7)	/* Read */
> #define HF_CMD_CA46		BIT(6)	/* Register space */
> #define HF_CMD_CA45		BIT(5)	/* Linear burst */
> 
> #define HF_CMD_READ_REG		(HF_CMD_CA47 | HF_CMD_CA46)

This will come into play for HyperRAM, not a care about for HyperFlash

> #define HF_CMD_READ_MEM		HF_CMD_CA47

For HyperFlash, this would be the only bit that needs to be set for read
along with HF_CMD_CA45 for linear burst.

> #define HF_CMD_WRITE_REG	HF_CMD_CA46
> #define HF_CMD_WRITE_MEM	0
> 
> MBR, Sergei
> 

-- 
Regards
Vignesh