Re: [PATCH] scsi: libsas: Grab the host lock in sas_ata_device_link_abort()

[Damien Le Moal <damien.lemoal@opensource.wdc.com>]

On 2022/12/21 18:44, Niklas Cassel wrote:
> On Wed, Dec 21, 2022 at 05:31:59PM +0900, Damien Le Moal wrote:
>>>
>>> What about the interrupt handler such as ahci_error_intr()? I didn't see
>>> the callers hold the port lock too. Do they need the port lock?
>>
>> It looks like it is missing for ahci_thunderx_irq_handler() but that one
>> takes the host lock. Same for xgene_ahci_irq_intr(), again no port lock
>> but host lock taken. And again for ahci_single_level_irq_intr() for the
>> non MSI case. For modern MSI adapters, the port lock is taken in
>>
>> For other cases, ahci_multi_irqs_intr_hard) takes the port lock.
>>
>> So it looks like ahci_port_intr() needs to take the lock and some
>> cleanups overall (the host lock should not be necessary in the command
>> path. But nobody seems to have issues with the "bad" cases... Probably
>> because they are not mainstream adapters.
>>
>> Definitely some work needed here.
> 
> ahci_multi_irqs_intr_hard() takes the ap->lock before calling
> ahci_handle_port_interrupt(), which calls ahci_port_intr(),
> so I don't think there is any work needed for multi IRQ AHCI.

Yes. I tried to say that above.

> 
> For ahci_single_level_irq_intr() the host lock is taken before
> calling ahci_handle_port_intr(), so I don't see why we need any
> extra work for single IRQ AHCI.
> 
> 
> Remember, while the default is that:
> 	ap->lock = &host->lock;

Ah ! Yes ! good point ! So there are no issues anywhere. This is only confusing
because of the comments I think.

> see:
> https://github.com/torvalds/linux/blob/v6.1/drivers/ata/libata-core.c#L5305
> 
> In case of MULTI MSI, the ap->lock is using its own lock:
> https://github.com/torvalds/linux/blob/v6.1/drivers/ata/libahci.c#L2460
> 
> 
> So what is it that needs to be fixed for AHCI?
> 
> I haven't looked at ahci_thunderx_irq_handler() and xgene_ahci_irq_intr()
> so I can't speak for these.

These are not multi-irq and use the &host->lock directly, which is the same as
taking the ap->lock. We could clean that up for consistency and always use
ap->lock. But not a big deal. No bugs, just a little confusing with a simple
reading of the code.

> 
> 
> Kind regards,
> Niklas

-- 
Damien Le Moal
Western Digital Research