Message-ID: <5831286b-e3d0-4b87-9c5c-dbcb420d1b67@gmail.com>
Date:   Mon, 16 Oct 2023 14:46:03 +0200
From:   Milan Broz <gmazyland@...il.com>
To:     Damien Le Moal <dlemoal@...nel.org>, Christoph Hellwig <hch@....de>
Cc:     linux-scsi@...r.kernel.org, jejb@...ux.ibm.com,
        martin.petersen@...cle.com, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] scsi: use ATA-12 pass-thru for OPAL as fallback

On 10/16/23 13:54, Damien Le Moal wrote:
> On 10/16/23 16:24, Milan Broz wrote:
>> On 10/16/23 09:05, Christoph Hellwig wrote:
>>> On Mon, Oct 16, 2023 at 09:02:11AM +0200, Milan Broz wrote:
>>>> All common USB/SATA or USB/NVMe adapters I tested need this patch.
>>>>
>>>> In short, these steps are run for OPAL support check:
>>>>     1) Storage driver enables security driver flag (security_supported).
>>>>        USB-attached storage drivers will enable it in a separate patchset.
>>>>        SCSI and NNVMe drivers do it already. If the flag is not enabled,
>>>>        no following steps are run, and OPAL remains disabled.
>>>>     2) SCSI device enumerates SECURITY IN/OUT command support. If detected,
>>>>        SECURITY ON/OUT wrapper is used (as in the current code).
>>>>        If not, new ATA-12 pass-thru wrapper is used instead.
>>>>     3) SED OPAL code tries OPAL discovery command for the device.
>>>>        If it receives a correct reply, OPAL is enabled for the device.
>>>>        If SCSI SECURITY or ATA-12 command with discovery command is rejected,
>>>>        OPAL remains disabled.
>>>>
>>>> Note, USB attached storage needs an additional patchset sent separately
>>>> as requested by USB driver maintainers (it contains required changes
>>>> related to USB quirk processing).
>>>
>>> This just feels wrong.  These adapters are broken if they can't
>>> translated, and we should not put ATA command submission into
>>> sd.c.
>>
>> I think it is blocked in USB layer as not running command enumeration,
>> SCSI SECURITY will be never sent to the adapter through USB.
>>
>> I understand the problem, but if you configure OPAL from userspace, ATA-12 is sent
>> to these devices already - so why kernel cannot use it too?
>>
>>>
>>>> +	cdb[0] = ATA_12;
>>>> +	cdb[1] = (send ? 5 /* ATA_PROTOCOL_PIO_DATA_IN */ : 4 /* ATA_PROTOCOL_PIO_DATA_OUT */) << 1;
>>>> +	cdb[2] = 2 /* t_length */ | (1 << 2) /* byt_blok */ | ((send ?  0 : 1) << 3) /* t_dir */;
>>>> +	cdb[3] = secp;
>>>> +	put_unaligned_le16(len / 512, &cdb[4]);
>>>> +	put_unaligned_le16(spsp, &cdb[6]);
>>>> +	cdb[9] = send ? 0x5e /* ATA_CMD_TRUSTED_SND */: 0x5c /* ATA_CMD_TRUSTED_RCV */;
>>>
>>>
>>> Also avoid all these crazy long lines, and please use the actual
>>> constants.  Using a good old if/else is actually a very good way to
>>> structure the code in a somewhat readable way.
>>
>> Sure, I was trying to no add additional includes that will mess this up, I'll reformat it if needed.
>>
>> Otherwise, this wrapper is exactly what is used is sedutils and also in our test utility
>> that tries to work with OPAL commands directly
>> https://github.com/mbroz/opal-toolset
>>
>>>
>>>> +		if (sdkp->security)
>>>> +		    sdkp->opal_dev = init_opal_dev(sdkp, &sd_sec_submit);
>>>> +		else
>>>> +		    sdkp->opal_dev = init_opal_dev(sdkp, &sd_ata12_submit);
>>>
>>> Messed up indentation here.
>>
>> sorry, my bad, I hate such formatting myself and missed it here :-)
>>    
>>> besides the fact that the statement is fundamentally wrong and you'll
>>> start sending ATA command to random devices.
>>
>> So what do you suggest? As I said, this exactly happen if you configure it from userspace.
>>
>> Can this be somehow limited? I did not find and way how to do it.
> The translation of SECURITY IN/OUT commands should go into usb uas.c, at the
> very least. And even having it there is not great in my opinion. If the adapter
> does not support Opal, don't use that feature, or use it only to lock/unlock the
> drive from user space with passthrough.

I was resisting to support OPAL hw for long time, but once we decided to add it as
an additional (and optional) layer for LUKS, I would like it to be supported also
for external drives (if technically possible).

The problem is that we (for simplicity) decided to use kernel SED-ioctl interface that
internally wraps OPAL command to SCSI SECURITY command only. It means, that all devices
that can use ATA-12 just cannot work with this kernel interface (unlike userspace which
can decide which wrapper to use).

And IMO it is not correct - if it was designed only for some servers with directly connected
devices, then it is really not generic OPAL support. It should work for any hw that supports it.

For USB, it actually works quite nice with the patch (ignoring usual bugs in firmware).

> 
> Note that nowhere in your patch do you test if you are talking to an ATA device.

Yes, I know. I expected the command to be rejected if not supported.

> This can be done by testing for the existence of VPD page 89h. See
> scsi_cdl_enable() in drivers/scsi/scsi.c for an example where we had to check
> for that. But also note that we do not issue ATA commands based on that test. We
> keep issuing SCSI commands and libata takes care of the translation. uas does
> not use libata though, so if translation is needed, do it there.

So, you mean translate SCSI SECURITY to ATA-12 inside USB storage drivers?

(There are actually two places, UAS driver and then SCSI glue for mass-storage -
unfortunately, we need both.)
  
> But I have the same opinion as Christoph: working around USB adapters lack of
> support for a feature with passthrough commands issued from the kernel is really
> not ideal.

Well, I have several adapters and many OPAL drives, none works with SCSI commands
if connected through USB. Partially it is missing support in USB layer, but the rest
is mess in hw. I know it is **** but that's how it is; people have these and want
to use it (including myself).

IMO it is quite similar to discard/TRIM support...

Thanks,
Milan

Powered by blists - more mailing lists