Message-ID: <687afe8d-87ca-4699-c886-4a6c03a47132@broadcom.com>
Date:   Sun, 8 Oct 2023 16:46:11 -0700
From:   William Zhang <william.zhang@...adcom.com>
To:     Miquel Raynal <miquel.raynal@...tlin.com>
Cc:     dregan@...l.com, bcm-kernel-feedback-list@...adcom.com,
        linux-mtd@...ts.infradead.org, f.fainelli@...il.com,
        rafal@...ecki.pl, joel.peshkin@...adcom.com,
        computersforpeace@...il.com, dan.beygelman@...adcom.com,
        frieder.schrempf@...tron.de, linux-kernel@...r.kernel.org,
        vigneshr@...com, richard@....at, bbrezillon@...nel.org,
        kdasu.kdev@...il.com
Subject: Re: [PATCH v2] mtd: rawnand: brcmnand: Initial exec_op implementation



On 10/06/2023 12:42 AM, Miquel Raynal wrote:
> Hi William,
> 
> william.zhang@...adcom.com wrote on Thu, 5 Oct 2023 17:42:21 -0700:
> 
>> Hi Miquel,
>>
>> On 10/03/2023 09:47 PM, William Zhang wrote:
>>>
>>>
>>> On 10/03/2023 03:55 PM, Miquel Raynal wrote:
>>>> Hi William,
>>>>
>>>> william.zhang@...adcom.com wrote on Tue, 3 Oct 2023 11:46:25 -0700:
>>>>   
>>>>> Hi Miquel,
>>>>>
>>>>> On 10/03/2023 02:28 AM, Miquel Raynal wrote:
>>>>>> Hi William,
>>>>>>
>>>>>> william.zhang@...adcom.com wrote on Mon, 2 Oct 2023 12:57:01 -0700:
>>>>>>> Hi Miquel,
>>>>>>>
>>>>>>> On 10/02/2023 05:35 AM, Miquel Raynal wrote:
>>>>>>>> Hi David,
>>>>>>>>
>>>>>>>> dregan@...l.com wrote on Sat, 30 Sep 2023 03:57:35 +0200:
>>>>>>>>      >>>> Initial exec_op implementation for Broadcom STB, >>>>>> Broadband and iProc SoC
>>>>>>>>> This adds exec_op and removes the legacy interface.
>>>>>>>>>
>>>>>>>>> Signed-off-by: David Regan <dregan@...l.com>
>>>>>>>>> Reviewed-by: William Zhang <william.zhang@...adcom.com>
>>>>>>>>>
>>>>>>>>> ---
>>>>>>>>>     >>>
>>>>>>>> ...
>>>>>>>>      >>>> +static int brcmnand_parser_exec_matched_op(struct >>>>>> nand_chip *chip,
>>>>>>>>> +                     const struct nand_subop *subop)
>>>>>>>>> +{
>>>>>>>>> +    struct brcmnand_host *host = nand_get_controller_data(chip);
>>>>>>>>> +    struct brcmnand_controller *ctrl = host->ctrl;
>>>>>>>>> +    struct mtd_info *mtd = nand_to_mtd(chip);
>>>>>>>>> +    const struct nand_op_instr *instr = &subop->instrs[0];
>>>>>>>>> +    unsigned int i;
>>>>>>>>> +    int ret = 0;
>>>>>>>>> +
>>>>>>>>> +    for (i = 0; i < subop->ninstrs; i++) {
>>>>>>>>> +        instr = &subop->instrs[i];
>>>>>>>>> +
>>>>>>>>> +        if ((instr->type == NAND_OP_CMD_INSTR) &&
>>>>>>>>> +            (instr->ctx.cmd.opcode == NAND_CMD_STATUS))
>>>>>>>>> +            ctrl->status_cmd = 1;
>>>>>>>>> +        else if (ctrl->status_cmd && (instr->type == >>>>>>> NAND_OP_DATA_IN_INSTR)) {
>>>>>>>>> +            /*
>>>>>>>>> +             * need to fake the nand device write protect >>>>>>> because nand_base does a
>>>>>>>>> +             * nand_check_wp which calls nand_status_op >>>>>>> NAND_CMD_STATUS which checks
>>>>>>>>> +             * that the nand is not write protected before an >>>>>>> operation starts.
>>>>>>>>> +             * The problem with this is it's done outside >>>>>>> exec_op so the nand is
>>>>>>>>> +             * write protected and this check will fail until >>>>>>> the write or erase
>>>>>>>>> +             * or write back operation actually happens where we >>>>>>> turn off wp.
>>>>>>>>> +             */
>>>>>>>>> +            u8 *in;
>>>>>>>>> +
>>>>>>>>> +            ctrl->status_cmd = 0;
>>>>>>>>> +
>>>>>>>>> +            instr = &subop->instrs[i];
>>>>>>>>> +            in = instr->ctx.data.buf.in;
>>>>>>>>> +            in[0] = brcmnand_status(host) | NAND_STATUS_WP; /* >>>>>>> hide WP status */
>>>>>>>>
>>>>>>>> I don't understand why you are faking the WP bit. If it's set,
>>>>>>>> brcmnand_status() should return it and you should not care about >>>>>> it. If
>>>>>>>> it's not however, can you please give me the path used when we have
>>>>>>>> this issue? Either we need to modify the core or we need to provide
>>>>>>>> additional helpers in this driver to circumvent the faulty path.
>>>>>>>
>>>>>>> The reason we have to hide wp status for status command is because
>>>>>>> nand_base calls nand_check_wp at the very beginning of write and erase
>>>>>>> function. This applies to both exec_op path and legacy path. With
>>>>>>> Broadcom nand controller and most of our board design using the WP pin
>>>>>>> and have it asserted by default, the nand_check_wp function will fail
>>>>>>> and write/erase aborts.  This workaround has been there before this
>>>>>>> exec_op patch.
>>>>>>>
>>>>>>> I agree it is ugly and better to be addressed in the nand base >>>>> code. And
>>>>>>> I understand Broadcom's WP approach may sound a bit over cautious >>>>> but we
>>>>>>> want to make sure no spurious erase/write can happen under any
>>>>>>> circumstance except software explicitly want to write and erase.  >>>>> WP is
>>>>>>> standard nand chip pin and I think most the nand controller has that
>>>>>>> that pin in the design too but it is possible it is not used and
>>>>>>> bootloader can de-assert the pin and have a always-writable nand flash
>>>>>>> for linux. So maybe we can add nand controller dts option >>>>> "nand-use-wp".
>>>>>>> If this property exist and set to 1,  wp control is in use and nand
>>>>>>> driver need to control the pin on/ff as needed when doing write and
>>>>>>> erase function. Also nand base code should not call nand_check_wp when
>>>>>>> wp is in use. Then we can remove the faking WP status workaround.
>>>>>>>>      >>>> +        } else if (instr->type == NAND_OP_WAITRDY_INSTR) {
>>>>>>>>> +            ret = bcmnand_ctrl_poll_status(host, NAND_CTRL_RDY, >>>>>>> NAND_CTRL_RDY, 0);
>>>>>>>>> +            if (ctrl->wp_cmd) {
>>>>>>>>> +                ctrl->wp_cmd = 0;
>>>>>>>>> +                brcmnand_wp(mtd, 1);
>>>>>>>>
>>>>>>>> This ideally should disappear.
>>>>>>>>      >> Maybe we can have the destructive operation patch from Borris.
>>>>>>> Controller driver still need to assert/deassert the pin if it uses >>>>> nand
>>>>>>> wp feature but at least it does not need to guess the op code.
>>>>>>
>>>>>> Ah, yeah, I get it.
>>>>>>
>>>>>> Please be my guest, you can revive this patch series (might need light
>>>>>> tweaking, nothing big) and also take inspiration from it if necessary:
>>>>>> https://github.com/bbrezillon/linux/commit/e612e1f2c69a33ac5f2c91d13669f0f172d58717 >>>>
>>>>>> https://github.com/bbrezillon/linux/commit/4ec6f8d8d83f5aaca5d1877f02d48da96d41fcba >>>>
>>>>>> https://github.com/bbrezillon/linux/commit/11b4acffd761c4928652d7028d19fcd6f45e4696 >>>>
>>>>> Sure we will incorporate the destructive operation patch and provide a
>>>>> new revision.
>>>>>
>>>>> The WP status workaround will stay at least for this change. If you
>>>>> think my suggestion using a dts setting above is okay, we can provide a
>>>>> patch for that as well.  Or if you have any other idea or suggestion,
>>>>> we'd like to hear too.
>>>>
>>>> I thought this was not needed as Boris initial conversion did not need
>>>> it. The goal is to get rid of this workaround.
>>>> Boris' initial patch did remove that workaround but it will break the
>>> board that uses WP pin because the nand_check_wp run before the exec_op > and status returned is write-protected in the erase and write function.
>>> I explained that above and you can see the code here:
>>> https://elixir.bootlin.com/linux/v6.6-rc4/source/drivers/mtd/nand/raw/nand_base.c#L4599 > >
>>> I agree with your goal to remove this workaround and we have suggested
>>> one possible fix but we are also open to any other solution.
>>>    
>> We have integrated the destructive operation patch and are ready for the
>> v3.  If you don't think my proposal on the WP status fix is a good idea,
>> can we get this exce_op conversion patch series going first?  After all,
>> we don't modify the WP status handling behavior in this patch. We can
>> fix it in another patch whenever we agree on a solution.  Please let me
>> know and thanks a lot for all your comments and thoughts.
> 
> The NAND core has been a playground for coding horrors sometimes, and
> this ->exec_op() conversion is us the way to a cleaner and mastered
> approach, I am not willing to let something that obvious get in, I'm
> sorry. For you it's just a workaround, for me it means any change in
> the core will just break with this controller.
> 
> This is of course not against you or your work, perhaps I should
> emphasize that I strongly appreciate your efforts and, besides this
> workaround the code is clean.
> 
> The problem is that the WP pin can be used in two different ways:
> internally and externally. When it's used externally, you expect it
> to be deasserted before you start a destructive operation. When you use
> it internally, you expect it to be deasserted during the destructive
> operation.
> 
> The final solution needs to be approved by comparing with
> similar drivers which perform this internal procedure themselves
> as well. Maybe we could add a flag somewhere in the core's controller
> structure to tell the core not to perform these checks as we master the
> handling of the WP pin, telling the controller will handle it
> correctly as long as the destructive flag is passed.
> 
Thanks for the comments and I think we are on the same page. We will add
the flag in the core controller's and provide the updated patches.

> Thanks, Miquèl
> 

Download attachment "smime.p7s" of type "application/pkcs7-signature" (4212 bytes)

Powered by blists - more mailing lists