On 5/23/23 18:23, Jason Yan wrote:
> On 2023/5/23 16:29, John Garry wrote:
>> On 23/05/2023 09:04, Damien Le Moal wrote:
>>> For devices not attached to a port multiplier and managed directly by
>>> libata, the device number passed to ata_find_dev() must always be lower
>>> than the maximum number of devices returned by ata_link_max_devices().
>>> That is 1 for SATA devices or 2 for an IDE link with master+slave
>>> devices. This device number is the scsi device ID which matches these
>>> constraint as the ID are generated per port and so never exceed the
>>> link maximum.
>>>
>>> However, for libsas managed devices, scsi device IDs are assigned per
>>> scsi host, leading to device IDs for SATA devices that can be well in
>>> excess of libata per-link maximum number of devices. This results in
>>> ata_find_dev() always returning NULL for libsas managed devices except
>>> for the first device of the host with ID (device number) 0. This issue
>>> is visible by executing hdparm command, which fails:
>>>
>>> hdparm -i /dev/sdX
>>> /dev/sdX:
>>> HDIO_GET_IDENTITY failed: No message of desired type
>>>
>>> Fix this by rewriting ata_find_dev() to ignore the device number for
>>> non-pmp attached devices with a link with at most 1 device, that is SATA
>>> devices on SATA ports. For these, device number 0 is always used to
>>> return the correct ata_device struct of the port link. This change
>>> excludes IDE master/slave setups (maximum number of devices per link
>>> is 2) and port-multiplier attached devices. Also, to be consistant with
>>> the fact that scsi device IDs and channel numbers used as device numbers
>>> are both unsigned int, change the devno argument of ata_find_dev() to
>>> unsinged int.
>>>
>>> Reported-by: Xingui Yang <yangxingui@xxxxxxxxxx>
>>> Fixes: 41bda9c98035 ("libata-link: update hotplug to handle PMP links")
>>> Cc: stable@xxxxxxxxxxxxxxx
>>> Signed-off-by: Damien Le Moal <dlemoal@xxxxxxxxxx>
>>
>> Reviewed-by: John Garry <john.g.garry@xxxxxxxxxx>
>
> Hi Damien & John,
>
> I think we may missed something. What about if we do this:
>
> echo "scsi add-single-device 0 0 100 0" >/proc/scsi/scsi
>
> Then in ata_scsi_user_scan() we will return device "0 0 0 0 " and rescan
> this device, which is wrong?
I did some more digging into this. And I do not think there are any
issues (I tested and it works). The reason is that the "shost" passed to
ata_scsi_user_scan() corresponds to the scsi host for the ata_port of
the ata_device in libsas (dev->sata_dev.ap). It is not the scsi_host
representing the HBA itself (which has multiple devices).
So changing the devno to 0 has no effect and deleting+rescanning
particular devices leads to the correct IDs being used. E.g., on my
system, I have 4 drives connected to the pm80xx:
# lsscsi -g
[0:0:0:0] disk ATA WDC WUH721818AL W232 /dev/sdc /dev/sg3
[0:0:1:0] disk ATA WDC WUH721818AL WTW2 /dev/sdb /dev/sg1
[0:0:2:0] disk ATA WDC WUH722222AL Wf86 /dev/sdd /dev/sg2
[0:0:3:0] zbc ATA WDC WSH722020AL W803 /dev/sda /dev/sg0
If I remove the first 0 and 2:
# echo 1 > /sys/class/scsi_device/0:0:0:0/device/delete
# echo 1 > /sys/class/scsi_device/0:0:2:0/device/delete
# lsscsi -g
[0:0:1:0] disk ATA WDC WUH721818AL WTW2 /dev/sdb /dev/sg1
[0:0:3:0] zbc ATA WDC WSH722020AL W803 /dev/sda /dev/sg0
And then manually rescan in reverse order of the removal:
# echo "scsi add-single-device 0 0 2 0" > /proc/scsi/scsi
# lsscsi -g
[0:0:1:0] disk ATA WDC WUH721818AL WTW2 /dev/sdb /dev/sg1
[0:0:2:0] disk ATA WDC WUH722222AL Wf86 /dev/sdc /dev/sg2
[0:0:3:0] zbc ATA WDC WSH722020AL W803 /dev/sda /dev/sg0
# echo "scsi add-single-device 0 0 0 0" > /proc/scsi/scsi
# lsscsi -g
[0:0:0:0] disk ATA WDC WUH721818AL W232 /dev/sdd /dev/sg3
[0:0:1:0] disk ATA WDC WUH721818AL WTW2 /dev/sdb /dev/sg1
[0:0:2:0] disk ATA WDC WUH722222AL Wf86 /dev/sdc /dev/sg2
[0:0:3:0] zbc ATA WDC WSH722020AL W803 /dev/sda /dev/sg0
I get back all devices with the correct IDs.
Note that I tried John's suggestion as well using an ata_for_each_dev()
loop. That does work well for libsas as we do have ata_dev->sdev set to
the scsi device already when scanning, but does not work for AHCI as we
do not. I will keep doing some more tests with v3 (and correct typos and
suggested) but I think this is all good.
--
Damien Le Moal
Western Digital Research
