Re: [PATCH] pata_parport: add driver (PARIDE replacement)

[Damien Le Moal]

On 3/16/22 20:28, Ondrej Zary wrote:
> On Wednesday 16 March 2022, Sergey Shtylyov wrote:
>> Hello!
>>
>> On 3/14/22 12:19 AM, Ondrej Zary wrote:
>>
>> [...]
>>>>> The pata_parport is a libata-based replacement of the old PARIDE
>>>>> subsystem - driver for parallel port IDE devices.
>>>>> It uses the original paride low-level protocol drivers but does not
>>>>> need the high-level drivers (pd, pcd, pf, pt, pg). The IDE devices
>>>>> behind parallel port adapters are handled by the ATA layer.
>>>>>
>>>>> This will allow paride and its high-level drivers to be removed.
>>>>>
>>>>> paride and pata_parport are mutually exclusive because the compiled
>>>>> protocol drivers are incompatible.
>>>>>
>>>>> Tested with Imation SuperDisk LS-120 and HP C4381A (both use EPAT
>>>>> chip).
>>>>>
>>>>> Note: EPP-32 mode is buggy in EPAT - and also in all other protocol
>>>>> drivers - they don't handle non-multiple-of-4 block transfers
>>>>> correctly. This causes problems with LS-120 drive.
>>>>> There is also another bug in EPAT: EPP modes don't work unless a 4-bit
>>>>> or 8-bit mode is used first (probably some initialization missing?).
>>>>> Once the device is initialized, EPP works until power cycle.
>>>>>
>>>>> So after device power on, you have to:
>>>>> echo "parport0 epat 0" >/sys/bus/pata_parport/new_device
>>>>> echo pata_parport.0 >/sys/bus/pata_parport/delete_device
>>>>> echo "parport0 epat 4" >/sys/bus/pata_parport/new_device
>>>>> (autoprobe will initialize correctly as it tries the slowest modes
>>>>> first but you'll get the broken EPP-32 mode)
>>>>>
>>>>> Signed-off-by: Ondrej Zary <linux@xxxxxxx>
>>>> [...]
>>>>> diff --git a/Documentation/admin-guide/blockdev/paride.rst b/Documentation/admin-guide/blockdev/paride.rst
>>>>> index e1ce90af602a..e431a1ef41eb 100644
>>>>> --- a/Documentation/admin-guide/blockdev/paride.rst
>>>>> +++ b/Documentation/admin-guide/blockdev/paride.rst
>>>> [...]
>>>>> diff --git a/drivers/ata/pata_parport.c b/drivers/ata/pata_parport.c
>>>>> new file mode 100644
>>>>> index 000000000000..783764626a27
>>>>> --- /dev/null
>>>>> +++ b/drivers/ata/pata_parport.c
>>>>> @@ -0,0 +1,819 @@
>>>> [...]
>>>>> +static void pata_parport_lost_interrupt(struct ata_port *ap)
>>>>> +{
>>>>> +	u8 status;
>>>>> +	struct ata_queued_cmd *qc;
>>>>> +
>>>>> +	/* Only one outstanding command per SFF channel */
>>>>> +	qc = ata_qc_from_tag(ap, ap->link.active_tag);
>>>>> +	/* We cannot lose an interrupt on a non-existent or polled command */
>>>>> +	if (!qc || qc->tf.flags & ATA_TFLAG_POLLING)
>>>>> +		return;
>>>>> +	/*
>>>>> +	 * See if the controller thinks it is still busy - if so the command
>>>>> +	 * isn't a lost IRQ but is still in progress
>>>>> +	 */
>>>>> +	status = pata_parport_check_altstatus(ap);
>>>>> +	if (status & ATA_BUSY)
>>>>> +		return;
>>>>> +
>>>>> +	/*
>>>>> +	 * There was a command running, we are no longer busy and we have
>>>>> +	 * no interrupt.
>>>>> +	 */
>>>>> +	ata_port_warn(ap, "lost interrupt (Status 0x%x)\n", status);
>>>>> +	/* Run the host interrupt logic as if the interrupt had not been lost */
>>>>> +	ata_sff_port_intr(ap, qc);
>>>>> +}
>>>>
>>>>    As I said, ata_sff_lost_interrupt() could be used instead...
>>>
>>> It couldn't be used because it calls ata_sff_altstatus().
>>
>>    And? That one used to call the sff_check_altstatus() method (which you define)
>> even before my patch:
>>
>> https://git.kernel.org/pub/scm/linux/kernel/git/dlemoal/libata.git/commit/?h=for-next&id=03c0e84f9c1e166d57d06b04497e11205f48e9a8
> 
> OK, I was probably confused by ata_sff_check_status which uses ioread directly.
> 
>> [...]
>>>>> diff --git a/include/linux/pata_parport.h b/include/linux/pata_parport.h
>>>>> new file mode 100644
>>>>> index 000000000000..f1ba57bb319c
>>>>> --- /dev/null
>>>>> +++ b/include/linux/pata_parport.h
>>>>> @@ -0,0 +1,108 @@
>> [...]
>>>>> +static inline u16 pi_swab16(char *b, int k)
>>>>> +{
>>>>> +	union { u16 u; char t[2]; } r;
>>>>> +
>>>>> +	r.t[0] = b[2 * k + 1]; r.t[1] = b[2 * k];
>>>>> +	return r.u;
>>>>> +}
>>>>> +
>>>>> +static inline u32 pi_swab32(char *b, int k)
>>>>> +{
>>>>> +	union { u32 u; char f[4]; } r;
>>>>> +
>>>>> +	r.f[0] = b[4 * k + 1]; r.f[1] = b[4 * k];
>>>>> +	r.f[2] = b[4 * k + 3]; r.f[3] = b[4 * k + 2];
>>>>> +	return r.u;
>>>>
>>>>    Hey, I was serious about swab{16|32}p()! Please don't use home grown byte
>>>> swapping...
>>>
>>> This crap comes from old paride.h and we can't get rid of it without touching the protocol drivers
>>
>>    I don't argue about the *inline*s themselves, just about the ineffective code inside them.
>>
>>> (comm.c and kbic.c). Maybe use something like:
>>>
>>> #define pi_swab16(char *b, int k) 	swab16p((u16 *)&b[2 * k])
>>
>>> but I'm not sure it's equivalent on a big-endian machine.
>>
>>    These functions are endian-agnostic -- they swap always.
>>    If you only need to swab the bytes on big-endian machines, you should use cpu_to_le*() and/or
>> le*_to_cpu()...
> 
> swab16 swaps always but pi_swab16 does not on big-endian. It's probably a bug but doing the correct thing by accident. Other protocol drivers completely ignore endianness, probably because PARIDE was meant for x86 only.

Fix that. ATA/IDE uses little endian. So all command & replies fields
should be handled with put_unaligned_lexx()/get_unaligned_lexx(), or
cpu_to_lexx() and lexx_to_cpu().

> 
>> [...]
>>
>> MBR, Sergey
>>
> 
> 
> 


-- 
Damien Le Moal
Western Digital Research