[Linux-parport] [PATCH] pata_parport: add driver (PARIDE replacement)
Damien Le Moal
damien.lemoal at opensource.wdc.com
Mon Dec 12 15:07:31 PST 2022
On 12/13/22 07:55, Ondrej Zary wrote:
> On Wednesday 16 November 2022 02:30:46 Damien Le Moal wrote:
>> On 2022/11/15 23:56, Ondrej Zary wrote:
>>> On Tuesday 15 November 2022, Damien Le Moal wrote:
>>>> On 11/15/22 04:25, Ondrej Zary wrote:
>>>>> On Monday 14 November 2022 09:03:28 Damien Le Moal wrote:
>>>>>> On 11/14/22 16:53, Ondrej Zary wrote:
>>>>>>> On Monday 14 November 2022, Damien Le Moal wrote:
>>>>>>>> On 11/12/22 20:17, Ondrej Zary wrote:
>>>>>>>>> On Wednesday 19 October 2022 09:34:31 Christoph Hellwig wrote:
>>>>>>>>>> It's been a while - did you get a chance to make some progress on
>>>>>>>>>> this? Do you need any help to unblock you?
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Sorry again, I'm back now. Trying to fix locking problems.
>>>>>>>>> Added this to each function for analysis how the functions are called wrt.
>>>>>>>>> locking:
>>>>>>>>>
>>>>>>>>> printk("%s, locked=%d\n", __FUNCTION__, spin_is_locked(ap->lock));
>>>>>>>>
>>>>>>>> Do you have your code somewhere that we can look at ?
>>>>>>>
>>>>>>> This is the current version with debug printks. I've also added dump_stack()
>>>>>>> to find out the code path but haven't analyzed the output yet.
>>>>>>
>>>>>> Can you send a proper patch ? Or a link to a git tree ? That is easier to
>>>>>> handle than pasted code in an email...
>>>>>
>>>>> Patch against what? I don't have a git server.
>>>>
>>>> patch against current 6.1-rc, or against an older kernel should be OK too.
>>>> But please "git send-email" a patch, or push your dev tree to github ?
>>>>
>>>>> I've done some call trace analysis. These code paths are calling
>>>>> pata_parport functions with ap->lock locked during init.
>>>>>
>>>>> Comm: kworker, Workqueue: ata_sff ata_sff_pio_task
>>>>> ata_sff_hsm_move -> ata_pio_sectors-> ata_sff_altstatus -> pata_parport_tf_read -> pata_parport_check_altstatus
>>>>> ata_sff_hsm_move -> ata_sff_altstatus -> pata_parport_tf_read -> pata_parport_check_altstatus
>>>>> ata_sff_pio_task -> ata_sff_busy_wait -> pata_parport_check_status
>>>>> ata_sff_hsm_move -> ata_wait_idle -> ata_sff_busy_wait -> pata_parport_check_status
>>>>> ata_sff_hsm_move -> ata_hsm_qc_complete -> ata_sff_irq_on -> ata_wait_idle -> ata_sff_busy_wait -> pata_parport_check_status
>>>>> ata_sff_pio_task -> ata_sff_hsm_move -> ata_pio_sectors -> ata_pio_sector -> ata_pio_xfer -> pata_parport_data_xfer
>>>>> ata_sff_pio_task -> ata_sff_hsm_move -> pata_parport_data_xfer
>>>>> ata_sff_pio_task -> ata_sff_hsm_move -> pata_parport_tf_read
>>>>> ata_sff_hsm_move -> ata_hsm_qc_complete -> ata_qc_complete -> fill_result_tf -> ata_sff_qc_fill_rtf -> pata_parport_tf_read
>>>>> ata_sff_hsm_move -> ata_pio_sectors -> ata_sff_altstatus -> pata_parport_check_altstatus
>>>>> ata_sff_hsm_move -> ata_sff_altstatus -> pata_parport_check_altstatus
>>>>>
>>>>> Comm: modprobe
>>>>> ata_host_start -> ata_eh_freeze_port -> ata_sff_freeze -> pata_parport_check_status
>>>>>
>>>>> Comm: scsi_eh_4
>>>>> ata_eh_recover -> ata_eh_reset -> ata_eh_thaw_port -> ata_sff_thaw -> ata_sff_irq_on -> ata_wait_idle -> ata_sff_busy_wait -> pata_parport_check_status
>>>>> ata_eh_reset -> ata_eh_freeze_port -> ata_sff_freeze -> pata_parport_check_status
>>>>> ata_scsi_error -> ata_scsi_port_error_handler -> ata_port_freeze -> ata_sff_freeze -> pata_parport_check_status
>>>>> ata_sff_error_handler -> pata_parport_drain_fifo -> pata_parport_check_status
>>>>
>>>> What exactly are the issues you are having with ap->lock ? It looks like
>>>> you have done a lot of analysis of the code, but without any context about
>>>> the problem, I do not understand what I am looking at.
>>>>
>>>
>>> The problem is that pi_connect() can sleep because it calls
>>> parport_claim_or_block(). And any access (even reading ATA status register)
>>> requires pi_connect.
>>
>> OK. Let me have a look.
>>
>
> The locking problems seem not to be easily solvable. Maybe a hack that grabs
> the parport before registering ata interface (and keeps it until the
> interface is disabled) will help? That will prevent multiple chained devices
> on one parport from working but can get pata_parport moving.
I have been very busy and had no time to look at this. Sorry about that. I
will have a look before year-end break.
--
Damien Le Moal
Western Digital Research
More information about the Linux-parport
mailing list